NADI 2026 is the seventh edition of the Nuanced Arabic Dialect Identification (NADI) shared task series and the second dedicated to multidialectal Arabic speech processing. This edition comprises five tasks and eight subtasks spanning Automatic Speech Recognition (ASR), Spoken Dialect Identification (SDID), Text-to-Speech (TTS), Spoken Language Translation (SLT), and Spoken Language Understanding (SLU). NADI 2026 emphasizes realistic evaluation through low-bandwidth, mixed-dialect, code-switched, out-of-domain, and zero-shot settings, while introducing TTS, SLT, and SLU to the series for the first time. The shared task attracted 21 participating teams from at least 13 countries, with 48 test-phase submissions, and 14 submitted system-description papers. Results show that out-of-domain generalization remain a bottleneck and highlight the effectiveness of recent Arabic-specialized speech models, multimodal dialect identification, and ensemble methods. Overall, NADI 2026 provides a broader and more challenging benchmark for robust Arabic dialect speech processing.

This paper describes our submission to NADI 2026 Subtask 1.1, speech recognition for eight Arabic dialects under noisy conditions. Our system adapts whisper-large-v3-turbo with one low-rank adapter per dialect, decodes with tuned beam search, and applies a normalization step that matches the spelling conventions of the reference transcriptions rather than applying generic orthographic rules. We train only on the shared-task data, which places our results well below those of systems built with large external corpora, and we report our numbers with that constraint in mind. Country-average WER falls from 0.804 for the zero-shot backbone to 0.6053 on the test set. We also report three observations that shaped the system: per-dialect adapters outperform a jointly trained adapter on all eight dialects even when the joint model is trained for twice as long; replacing beam search with temperature sampling in order to produce diverse hypotheses for rescoring costs more accuracy than the rescoring recovers; and because the official scorer normalizes hypothesis and reference identically, generic character-level normalization cannot help, whereas learning each dialect’s annotation convention from the training transcriptions does.

We present Aslema, our system for NADI 2026 Shared Task 5, which consists of two subtasks: intent recognition and slot filling. We evaluate four omni LLMs in a zero-shot setting and compare them with fine-tuned models. Our results show that fine-tuning consistently outperforms zero-shot inference. We further explore synthetic data augmentation by using an LLM to generate culturally grounded Tunisian Derja utterances, followed by voice cloning to generate synthetic speech. Incorporating this synthetic data improves performance on both tasks. Our final submitted system, based on Qwen3-Omni-30B and trained with a mixture of original and synthetic data, achieves 86.8% intent accuracy and 34.7 WER on the devtest split. On the official test set it ranks 1st in slot filling (59.5 CoER) and 4th among 5 teams in intent recognition (66.1% accuracy). We release our experimental scripts and the synthetic dataset to support further research.

We describe our submission to three NADI 2026 subtasks: Tunisian Arabic–French–English code-switched ASR (Task 1.3), dialectal Arabic speech translation, and 20-way Arabic dialect identification (ADI). For code-switched ASR, we finetune a 2B-parameter FastConformer encoder-decoder with synthetic Tunisian speech, cross-lingual data, waveform augmentation, and checkpoint averaging, achieving 14.12% WER and ranking first overall. For speech translation, finetuning Whisper Large-v3 yields strong in-domain results but poor generalization (BLEU 5.8 on blind test), illustrating overfitting under limited data. For ADI, a classifier on frozen pretrained ASR representations outperforms full/partial finetuning, reaching 42% accuracy across 20 dialects. We highlight synthetic augmentation, checkpoint averaging, and encoder freezing as effective low-resource strategies.

Two-phase Whisper fine-tuning (broad dialect pre-training + low-LR target annealing) reaches 57.0 WER on NADI 2026 Subtask 1.1, with 8 WER points from the schedule alone.

Wifaq is our entry to NADI 2026 Subtask 1.1, which evaluates country-level Arabic automatic speech recognition across eight dialects under noisy conditions. We fully fine-tune four 600M-parameter W2v-BERT 2.0 CTC models on nested mixtures of task and ex- ternal speech, decode them with a shared 5-gram KenLM, and combine their hypotheses by word voting; each run fits on one 16 GB NVIDIA RTX 4070 Ti SUPER. The submitted four-way ensemble ranked first among nine entries on the blind test with 43.23% WER and 20.0% CER. On a reused 800-utterance local holdout, it lowers WER from 46.12% for the strongest single model to 44.0%

Robust Arabic speech processing must accommodate dialectal variation, multilingual code-switching, and distribution shift. We present NAMAA Community’s systems for three NADI 2026 speech subtasks: Mixed-Dialect Arabic ASR, Code-Switched Arabic ASR, and 20-way Spoken Arabic Dialect Identification. For Mixed-Dialect ASR, an Arabic-specific Conformer–Transformer achieves 0.45 word error rate (WER) on the blind set, outperforming a development-selected normalization and consensus pipeline at 0.49. For Code-Switched ASR, two-stage LoRA adaptation of the Tunisian Whisper-derived FarukSTT checkpoint, with multilingual tokenization and unconstrained decoding, yields 20.38% WER and 8.06% character error rate (CER) on validation and 16.98% blind-test WER. For dialect identification, a four-model ensemble reaches 0.950 development accuracy but 0.31 on the out-of-domain blind set. Label-free class-spread and inter-model agreement diagnostics identify an outlying component; removing it improves blind accuracy to 0.36, and a refined three-model submission reaches 0.38. Across Subtasks 1.2 and 2, development-selected calibration did not reliably transfer to blind data, although the available comparisons do not isolate the causal effect of individual components.

Open-set spoken intent recognition for low-resource Arabic varieties must handle both automatic speech recognition (ASR) errors and intents absent from training. We describe CIS-OpenSLU, our NADI 2026 Task 5.1 system, which transcribes Tunisian Arabic speech with Whisper Turbo and classifies the resulting text using word- and character-level TF–IDF features with a LinearSVC. On ASR-transcribed validation audio, training the classifier on Whisper-generated transcripts rather than gold transcripts raises weighted F1 from 49.25% to 69.78%, showing that matching training and inference text distributions is crucial. We rank test utterances by the maximum SVM decision score and reject low-confidence examples as unknown; the final system achieves 71.54% weighted F1 and 72.50% accuracy, ranking second. Auxiliary audio and multi-modal experiments further show that strong unseen-intent detection in leave-one-intent-out simulation does not necessarily transfer to genuinely unseen test intents.

We describe the CodeZone Research Group system for NADI 2026 Shared Task 2, spo ken Arabic dialect identification over Modern Standard Arabic and 19 country-level dialects under cross-domain evaluation. Our system ensembles an ECAPA-TDNN language identi fier fine-tuned on the shared task data with a Whisper-large-v3 dialect classifier that we fur ther adapt using waveform- and spectrogram level augmentation, fusing their log-softmax scores. This ensemble reaches 94.35% accu racy on in-domain validation but initially re tained only 25.40% on the official blind test, a collapse that affected every participating team. We show that about a fifth of this gap is a cali bration failure rather than a modelling failure: discarding per-class bias terms fit in domain raised official accuracy from 0.2540 to 0.3918 and lowered the shared task’s average cost from 0.2077 to 0.1515, and a softening temperature reduced cost further to 0.1469; neither change touched either model. Our submission ranked 10th of 16 teams, and we show analytically why uniform additive score shifts cannot affect this metric while temperature scaling can.

We describe our submission to NADI-2026 Subtask 2, country-level spoken Arabic dialect identification on a blind test set from undisclosed sources. Because the target domain is unknown at development time, we select on out-of-domain accuracy rather than in-domain validation accuracy at every decision point, comparing three backbones under a common pooling-plus-linear head and training the selected Arabic ASR encoder on ADI20-micro mixed with Casablanca. The system reaches 91.54% in-domain, 73.10% on the held-out MADIS-5 corpus (scored five-way), and 52.51% on the blind test set (C_avg 0.11), placing fourth. The criterion helped, choosing an encoder that in-domain accuracy would have rejected by 32 points on a held-out corpus, a gap more than twenty times our measured seed variation, but it was not sufficient: of the 39.03-point in-domain to blind gap, at least 18.44 points are attributable to leaving the training domain. We trace part of the remainder to a corpus that entered the training mixture while still being used for checkpoint selection.

We describe Thakaa’s systems for NADI 2026 subtasks 1.1 and 1.3 and Task 2, covering noisy country-level Arabic automatic speech recognition (ASR), Tunisian code-switched ASR, and spoken dialect identification. The ASR systems combine heterogeneous connectionist temporal classification (CTC), encoder–decoder, and audio-LLM recognizers through ROVER, with country-aware routing for subtask 1.1 and a frame-rate-corrected CTC anchor for subtask 1.3. The dialect identifier fuses Ara-BEST-RQ acoustic scores with lexical evidence extracted from Cohere and Whisper transcripts by a partially fine-tuned MARBERTv2 model. Our official test scores were 50.41% WER on subtask 1.1, 14.53% WER on subtask 1.3, and 56.15% accuracy on Task 2, ranking third, third, and first, respectively. Across the three systems, the largest gains came from explicit control of alignment, temporal resolution, and evidence diversity rather than from model scale alone.

We describe the KAND CA systems for all five subtasks of the NADI 2026 shared task on multidialectal Arabic speech processing. For the three ASR subtasks, a single family of staged, fully fine-tuned Whisper large-v3 models — broad dialectal adaptation, then only task-matched gold data — is combined with task-specific levers: an mT5 spellingcorrection post-editor trained on the task references (Subtask 1.1, final country-average WER 0.5111), chained continue-fine-tuning through the Subtask 1.1 data before the Mixed- ASR dev split (Subtask 1.2, WER 0.48), and a fresh two-epoch schedule over a 46K-clip external gold Tunisian–French corpus with an accent-only French post-pass (Subtask 1.3, WER 15.21%). For spoken dialect identification (Task 2), where in-domain accuracy anti-predicted blind-test accuracy, our system fuses an architecturally distinct FastConformerencoder classifier, a decorrelated enhancedaudio view, and an ASR-transcript text route, reaching 0.5114 accuracy. For dialectal TTS (Subtask 3) we submitted 4,001 voice-cloned clips across ten dialects (UTMOS 2.72). We also report extensive negative results and release our models, correctors, and code publicly.

Speech conveys complementary information through its linguistic content and acoustic characteristics. Speech-related tasks can exploit either source of information or combine them to improve performance. In this paper, we present Lynx_Id and Lynx_In, our systems for the NADI 2026 Spoken Dialect Identification and Intent Recognition tasks, respectively. Both systems build on an automatic speech recognition (ASR) model. Lynx_Id is a late-fusion system for dialect identification. It combines internal representations from the ASR encoder with predictions derived from the ASR transcript. Lynx_In is a two-branch text-based system that uses ASR transcriptions for intent recognition. One branch predicts the intent among the known classes, while the second acts as a gate to detect out-of-scope utterances. Lynx_Id achieved 53.53% accuracy and ranked third in the Spoken Dialect Identification task, while Lynx_In achieved a weighted F1 score of 76.86% and ranked first in the Intent Recognition task.

This paper describes the Resonate team’s submission to the NADI 2026 shared task on multidialectal Arabic speech processing. We participated in the robust dialectal Arabic ASR (Subtask 1.1) and spoken Arabic dialect identification (Task 2) subtasks, adding no external labelled speech. Our best single ASR model is a fully fine-tuned Cohere Transcribe Arabic, a 2.07B-parameter Arabic speech model, trained on 239.6K rows derived entirely from the 12.8K official clips through voice conversion, in-corpus noise augmentation and short-clip re-segmentation. Our submitted entry combines nine checkpoints of four sibling runs with a pivot-anchored ROVER vote and a per-dialect orthographic canonicaliser, obtaining 50.29% WER and 24.43% CER on the official test set and ranking second of nine teams. Our dialect identification system is a fine-tuned WavLM Base+ encoder with context-aware multi-head factorized attentive pooling under an angular-margin objective, which obtained 94% accuracy on the validation set and 41% on the out-of-domain test set. The ASR comparisons support augmenting scarce dialectal transcripts with acoustic variation, while ADI highlights the need for evaluation across domains. Code and reproducibility resources for both tasks are available at https://github.com/ResonateResearch/resonate-nadi-2026.

We describe the Itgan systems for the three ASR subtasks of NADI 2026, namely robust country-level ASR (1.1), mixed-dialect ASR (1.2), and Tunisian code-switched ASR (1.3). All three share one recipe, Whisper adapted with LoRA on consumer GPUs, and each was carried by a different addition to it. On 1.1, where the dialect label is given at test time, per-dialect specialists continued from a pooled adapter gave the largest gain, and the submitted system reached 57.1% country-average WER. A post-evaluation linear probe on frozen encoder features routes utterances without the label and recovers 44% of what oracle routing gives. On 1.2 the choice of base model mattered more than adapter capacity, and system combination helped only once we added a decorrelated member, reaching 46.7% WER. On 1.3 our system placed second at 14.49% WER with the lowest CER among the leading submissions, 5.38%. Its last 0.60 WER points came without further training, mostly from an exact weight-space average of independently trained runs, with ROVER voting adding the remainder. Every comparison carries a paired-bootstrap test, and we report eight directions that did not work.

We present the results and findings of the BAREC Shared Task 2026, the second shared task on Arabic readability assessment, organized as part of the Fourth Arabic Natural Language Processing Conference (ArabicNLP 2026). The shared task focuses on automatic readability assessment using the BAREC Corpus, addressing fine-grained classification across 19 readability levels. It comprises two subtasks —sentence-level classification and document-level classification— and three tracks: (1) the Strict Track, which permits only the BAREC Corpus; (2) the Constrained Track, which permits the BAREC Corpus, the SAMER Corpus, and the SAMER Lexicon; and (3) the Open Track, which permits the use of any external resources. A total of 14 teams from 12 countries registered for the task, of which 10 submitted system description papers. All participating teams competed only in the sentence-level classification subtask. The top-performing system achieved a quadratic weighted kappa (QWK) of 85.3.

This paper describes the system submitted by team HP for the BAREC Shared Task 2026 on Arabic readability assessment. Driven by the Quadratic Weighted Kappa (QWK) eval- uation metric, which heavily penalizes dis- tant misclassifications, we reframe the 19- level task as an ordinal regression problem rather than standard multi-class classification. Our approach leverages pre-trained language models (AraBERTv02, ARBERT) optimized via Parameter-Efficient Fine-Tuning (LoRA) and trained with a COnsistent RAnk Log- its (CORAL) loss framework incorporating sample-level class weights. To ensure robust generalization, we consolidate the available datasets and implement a 5-fold stratified cross- validation pipeline. Finally, to optimize the QWK metric on the official blind test, we apply an 18-threshold Nelder-Mead optimizer cali- brated exclusively on blended Out-Of-Fold log- its, yielding a blind test QWK of 0.837. Our solution is released in the Github.

Team thylinao entered the Open and Strict tracks of the BAREC 2026 shared task on sentence-level Arabic readability with a deliberately plain system: a fourteen-member equal-weight ensemble of encoder fine-tunes, decoded to the 19-level scale by cutpoints interpolated halfway between validation-tuned thresholds and thresholds that match the blind score distribution to the training label prior. It scored 85.3 QWK on both tracks, first on both. This paper is about how those choices were measured. Three independent estimates place the noise floor of our offline evaluation between 0.09 and 0.44 QWK, as large as the margins that decided the final ranking. In three paired comparisons against the live board, only the smallest of the three offline estimates survived; the other two reversed in sign. We therefore proved that the Open and Strict tracks score identical gold, and made every decision worth a submission from a live board reading. Once the board showed the mean was capped, every candidate we probed came back a tie, bar one that we later withdrew as outside the rules. The full ledger, including four refuted hypotheses, is in the appendix.

This paper presents the Northwestern University in Qatar (NU-Q) submission to Task 1 of the BAREC Shared Task 2026 on fine-grained Arabic readability assessment. The task requires predicting the readability level of Arabic sentences across 19 ordered readability levels. Our system fine-tunes CAMeLBERT-Mix as a 19-class classifier using the official BAREC training data and cross-entropy optimization. We evaluate the proposed system using Quadratic Weighted Kappa (QWK), the official shared-task evaluation metric, together with classification accuracy. Our submitted system achieves a validation QWK of 0.7393 and an accuracy of 45.96%, indicating that pretrained Arabic language models provide a strong baseline for fine-grained Arabic readability assessment.

Readability assessment has played a vital role in education, yet most NLP work has focused on English and a few high-resource languages, leaving fine-grained Arabic readability comparatively unexplored. In the BAREC Shared Task 2026, we have addressed sentence-level classification into 19 ordinal readability levels under the Strict Track, fine-tuning four Arabic transformer encoders, MARBERT, CAMeLBERT MSA, AraELECTRA and AraBERTv2, on the official training split under multiple ordinal-aware losses: soft cross-entropy, squared Earth Mover’s Distance, direct ordinal regression and a CORAL-style cumulative-threshold loss. A validation-only greedy ensemble of all resulting models, combined with a learned per-level threshold post-processing step, has raised performance from a best single-model QWK of 0.8341 to a final validation QWK of 0.8507. On the official Blind Test leaderboard, our system has achieved a QWK of 0.848, securing 3rd place in the Strict Track

Fine-grained Arabic readability assessment aims to estimate how difficult an Arabic sentence is for its intended readers by assigning it to one of 19 ordered readability levels. This task is particularly challenging because of Arabic’s morphological complexity and the imbalanced distribution of examples across the levels. To address these challenges, we develop a system for the Sentence-Level Strict Track of the BAREC Shared Task 2026. Using only the official BAREC data, the system combines D3Tok-based morphological information, normalized surface text, and complementary features within a multi-seed ensemble of AraBERTv2-based regression models. Our submission achieved a QWK of 84.1% on the official Blind Test. This result indicates that the proposed combination provides competitive performance for fine-grained Arabic readability assessment.

We present the system of Team Skywalker for the Sentence-level Strict track of the BAREC Shared Task 2026, which requires predicting one of 19 ordered Arabic readability levels for each sentence. Our system integrates complementary public Arabic pretrained language-model (PLM) experts fine-tuned with classification, regression, and ordinal objectives. Their continuous scores, confidence estimates, and rank-based features are combined using cross-fitted LightGBM meta-models. We further incorporate a cumulative ordinal branch that models level-boundary probabilities and a normalized-text branch that improves robustness to Arabic orthographic variation. Finally, a QWK-aware decoder maps the fused continuous score to one of the 19 readability levels using monotonic thresholds selected exclusively on the public validation set. We use only the official training data for supervision and the public validation labels for model selection. In addition to the final fusion system, we report single-PLM baselines and controlled ablations. The final configuration achieves a QWK of 86.30 on the public validation set and 84.20 on the CodaBench blind evaluation, ranking fourth among the submitted systems.

Fine-grained Arabic readability assessment is challenging due to the language’s rich morphology and the ordered nature of readability labels. We investigate this problem in the sentence-level Strict Track of the BAREC Shared Task 2026, which uses a 19-level readability scale. We evaluate six controlled configurations spanning engineered linguistic and morphological features, TF-IDF/SVD lexical representations, contextual transformer encoders, multi-encoder fusion, output-level ensembling, and ordinal-aware training, followed by a final refinement using a single EmbeddingGemma encoder. Contextual transformer representations provide the largest improvement over classical feature-based baselines, while multi-encoder fusion and voting yield smaller additional gains. In contrast, word-repetition and adjacent-word-swap augmentation degrades performance, indicating that these perturbations do not reliably preserve readability labels. Our strongest configuration combines EmbeddingGemma with Gaussian soft labels, KL-divergence training, expected-value decoding, class weighting, and best-checkpoint selection, achieving 84.0 QWK on the test set and 82.4 QWK on the official blind test set. Overall, the results suggest that explicitly accounting for the ordinal structure of readability labels is a promising alternative to simply increasing architectural or ensemble complexity. We further summarize the resulting design insights into a practical framework for Arabic ordinal readability modeling.

We describe the ZHR systems for the two sentence-level tracks of the BAREC 2026 Shared Task on Arabic readability assessment: Strict, where training is restricted to the BAREC training split, and Open, which allows any public data. Each sentence receives one of 19 ordinal levels, and systems are ranked by quadratic weighted kappa (QWK). Both tracks use one recipe. Arabic encoders are fine-tuned under five ordinal objectives and averaged with weights from greedy selection, and the averaged score is discretized by 18 thresholds fitted to maximize QWK; every fitted quantity uses the development split. The official Strict system, an ensemble of ten train-split models, scores 85.0 QWK on the blind set. The official Open system adds pretraining on the silver-labeled BAREC-10M corpus and restricts selection to members whose early stopping never touched the calibration split; it scores 85.2. Both clear the official baseline (82.5) by 2.5 and 2.7 QWK while trailing it in exact accuracy, a trade-off that comes from the metric itself and that we measure directly. Negative results are part of the record: a catalog of 19 changes that did not help, each measured under a held-out protocol.

We present the Linguistic Security Institute (LSI) submission to the BAREC Shared Task 2026 (Task 1): predicting Arabic sentence readability on a 19-level ordinal scale. Our approach uses a logit-level ensemble of three Arabic pre-trained language models, with CAMeLBERT trained using an ordinal Earth Mover’s Distance loss and augmented by six SAMER Lexicon difficulty features. Controlled comparisons show that surface heuristics improve undertrained baselines but degrade converged models, and retrieval-augmented in-context learning yields no gains in this fine-grained setting. Our final ensemble achieves 81.7 QWK, ranking 1st among Constrained-track system submissions and 3rd among Open-track system submissions.

We describe Soloeval’s submission to the Sentence-Level Open Track of the BAREC Shared Task 2026 on fine-grained Arabic readability assessment. Rather than treating the 19-level readability scale as a scalar regression target, we fine-tune pretrained Arabic BERT encoders (MARBERTv2 and CAMeLBERT-Mix) with a soft-label cross-entropy objective whose target distribution is a Gaussian centered on the gold level, chosen to more directly reflect the Quadratic Weighted Kappa (QWK) evaluation metric’s quadratic distance penalty. We combine this with lightweight auxiliary surface features (sentence length, word count, diacritic density) and K-fold cross-validation. Our best single-model configuration, CAMeLBERT-Mix trained with F=3 folds, achieves a QWK of 80.6 on the BAREC 2026 blind test leaderboard for the Open track. We report our full hyperparameter search, discuss the gap between our dev- and test-set results, and outline what we would try next.

AraGenre 2026 is a shared task on hierarchical, definition-guided Arabic genre classification under limited supervision and substantial distribution shift. Systems assign each text a broad label and a specific label using an inventory of six broad branches and 74 defined final genres. The final benchmark contains 27,972 source-derived Arabic segments, spanning Modern Standard Arabic, Classical Arabic, and dialectal material, while the released labelled examples are small and primarily synthetic. Final specific classes are absent from the released labelled data, but their English names and definitions are provided. The task received 46 registrations and 373 submissions, with 17 ranked final entries and 11 system description papers. Thaka achieved the highest official Hierarchical Macro F1 at 73.52, followed by HP at 71.69 and NAMAA-1 at 70.13. Broad classification was consistently stronger than specific classification. The documented systems frequently combine label semantics with multilingual retrieval, LLM judging, pseudo-labelling, reranking, or calibration. We clarify the resource-driven taxonomy and inherited gold labels, distinguish combined distribution shift from isolated unseen-label effects, and provide systematic validation of the final evaluation data.

This paper presents the first investigation into the performance of the Conformer architecture for hierarchical text classification, specifically applied to predicting two-level Arabic text genres. The proposed system combines multilingual sentence embeddings, target label definition representations, a Conformer projection module, an auxiliary broad-genre classifier, synthetic data augmentation, and seed-based ensembling. Evaluated on the official shared task, the approach achieved a Hierarchical Macro-F1 of 0.4399, demonstrating strong high-level genre classification (Broad Macro-F1 of 0.6288) while highlighting ongoing challenges in distinguishing semantically close specific subcategories (Specific Macro-F1 of 0.2511).

This paper presents the system submitted by team HP for the AraGenre 2026 Shared Task on Hierarchical Definition-Guided Arabic Genre Classification. To address the low-resource constraint of the provided dataset (140 labeled training samples), we propose a two-stage pipeline combining LLM-based pseudo-labeling with supervised fine-tuning. In the first stage, we evaluate several models under a 5-fold cross-validation framework, including Gemini 3.5 Flash Lite (0.9413), Qwen2.5 7B (0.7203), LLaMA 3.1 8B (0.6380), and a multilingual NLI model (0.3099). Based on these results, Gemini 3.5 Flash Lite is selected to generate pseudo-labels for the unlabeled test set using a definition-guided few-shot prompt. In the second stage, MARBERTv2 and ARBERTv2 are fine-tuned on approximately 28,000 pseudo-labeled samples. Our experiments show that MARBERTv2 achieves the best standalone score of 0.7169 Hierarchical Macro F1, outperforming both ARBERTv2 (0.7118) and a soft voting ensemble (0.7165), suggesting that Arabic-specialized pretraining of MARBERTv2 is more effective than model ensembling for this task.

This paper describes the Northwestern University in Qatar submission to the AraGenre 2026 Shared Task on hierarchical Arabic genre classification. We formulate the task as a definition-guided dense retrieval problem by encoding Arabic documents and English genre definitions into a shared semantic space using the multilingual BGE-M3 model. Genre labels are assigned by retrieving the most semantically similar definition without task specific fine tuning. On the development set, the proposed approach achieved a Broad Macro F1 of 0.6125, a Specific Macro F1 of 0.6125, and a Hierarchical Macro F1 of 0.6125. These results show that a lightweight retrieval based framework is an effective approach for hierarchical Arabic genre classification.

AraGenre 2026 requires hierarchical classification of Arabic texts into evaluation-specific genres unseen during training, using supplied English definitions and taxonomy information. We present a transductive zero-shot system combining multilingual retrieval over multi-view label representations with per-genre z-normalization, calibrated Qwen3-8B candidate reranking, sibling-focused reranking, metric-aware macro-F1 decoding, and margin-gated shortlist widening. On the official 27,972-instance test set, the system achieves 0.5800 Hierarchical Macro F1, comprising 0.749 broad and 0.411 specific macro F1. Per-genre normalization provides the largest build-up improvement (+0.095), while an encoder-only diagnostic attributes +0.082 to LLM reranking. Calibrated log-probability fusion, multi-view labels, and decoding refinements provide further complementary gains.

We describe the MohannadHendi submission to the AraGenre 2026 shared task. The system casts hierarchical Arabic genre classification as cross-lingual retrieval between Arabic texts and English genre definitions, allowing it to consider labels unseen during training. We fine-tune multilingual E5-large on 140 texts paired with five deterministic views of their genre definitions. At inference, the nearest normalized definition embedding supplies both the broad and specific labels. The best local configuration achieved 0.9020 specific Macro F1. The submitted system obtained 0.4560 Hierarchical Macro F1 and ranked eighth among 18 final leaderboard entries. Its broad Macro F1 (0.6125) exceeded its specific Macro F1 (0.2995), showing that fine-grained genre discrimination remains the main difficulty.

Hierarchical, definition-guided genre classification places a short Arabic text under a broad genre and, within it, a fine-grained specific genre, using only a natural-language definition of each. We present NAMAA’s submission to AraGenre 2026. Fine-tuned bi-encoder classifiers reached 0.93 hierarchical macro F1 on development via data augmentation and cross-lingual pretraining, but collapsed to as low as 0.22 on the released test set, having specialized to a hand-engineered definition style that missed the test set’s larger, 74-genre taxonomy. We root-caused this gap and switched to zero-shot LLM classification, matching the best fine-tuned ensemble’s test score (0.6899 hierarchical F1) with no training; layering self-consistency sampling and disagreement resolution raised this further to 0.7013, ranking 3rd of 18 teams. Specific Macro F1 (0.5114) lags well behind Broad Macro F1 (0.8913), showing that most remaining errors concentrate in fine-grained genre choices rather than broad-category mistakes. Zero-shot LLM prompting generalizes across this definition and taxonomy shift where fine-tuned, definition-conditioned encoders do not. We release our code and system resources to support reproducibility and further research.[<https://github.com/NAMAA-ORG/Aragenre-Definition-Guided-Genre>].

We describe the GHADAI approach to the AraGenre 2026 shared task, which requires classifying Arabic text into broad and specific genre categories, including specific genres unseen during training. Our approach is a two-stage pipeline: a fine-tuned multilingual bi-encoder scores all 27,972 test sentences against the 74 candidate definitions, and a large language model reranks the top candidates per sentence using cyclic option rotation to mitigate positional bias. On the hidden test set, the bi-encoder achieves a hierarchical macro-F1 of 0.4213, which increases to 0.5006 with Fanar-9B reranking and to 0.6050 with Qwen-32B reranking over the top 15 candidates using five rotations. The final pipeline ranked fourth on the official leaderboard. These results demonstrate the effectiveness of combining definition-based retrieval with LLM reranking for generalization to unseen Arabic text genres.

We describe SynAnchor, Nasaq’s definition-guided system for AraGenre 2026. The task requires hierarchical Arabic genre classification under disjoint specific-genre inventories across training, development, and test, making generalization to unseen genres a central challenge. SynAnchor addresses this setting by generating Arabic anchor texts from genre definitions using GPT-4o and embedding them with BGE-M3 for similarity-based retrieval. High-confidence cases are classified directly, while ambiguous inputs are selectively adjudicated by GPT-5 among the three highest-ranked candidate genres. On the official test set of 27,972 texts spanning 74 unseen specific genres, SynAnchor achieves a Hierarchical Macro-F1 of 0.6006 on Codabench, compared with 0.2472 for the organizers’ definition-embedding baseline (approximately 2.4). Approximately 25% of test instances are resolved using retrieval alone, without LLM adjudication. The results suggest that combining definition-guided Arabic anchor synthesis with selective LLM reasoning can support hierarchical genre classification under changing label inventories.

We present our system SAGE: Sentence Analysis Genre Evaluation for the AraGenre shared task, which assigns Arabic sentences a hierarchical genre label from an English definition alone, with none of the 74 evaluation genres represented in the official training data. Since a definition is not itself an instance of its genre, we fine-tune the encoder generating the embeddings of both the sentences to classify and the candidate genre definitions and train a small adapter that refines those of the definitions. Studying three encoders and four datasets that were generated for training purposes, we find that the adapter’s training data matters more than either encoder choice or adapter hidden layer. Our best system, combining a fine-tuned encoder and an adapter, reached a Hierarchical Macro F1 score of 0.5278, placing seventh.

We present WathiqTeam’s system for the AraGenre 2026 Shared Task on hierarchical Arabic genre classification. The task requires assigning each Arabic text both one of six broad genres and one of 74 fine-grained specific genres organised beneath them. This setting is particularly challenging because many specific genres overlap in topic and vocabulary, while the hierarchy is highly asymmetric: the six broad genres contain substantially different numbers of specific children. We address these challenges with a broad-first retrieval architecture. Multilingual E5 compares Arabic texts with English genre definitions, an Arabic sentence encoder retrieves generated Arabic prototypes, and BM25 preserves exact lexical and structural cues. At the broad level, class-balanced pooling prevents parents with larger prototype pools from receiving more votes. After hard hierarchical gating, Weighted Reciprocal Rank Fusion combines the three independent rankings for the remaining specific labels. Qwen3 reranks only uncertain top-two broad decisions and is constrained by internal ensemble support. The submitted system achieved 0.4559 Hierarchical Macro F1, with 0.6250 Broad Macro F1 and 0.2869 Specific Macro F1. Component-oriented official runs show that class balancing and level-specific fusion produced most of the gain, while selective LLM reranking provided a small final improvement.

We report the first-place system in AraGenre 2026 (0.7352 Hierarchical Macro F1, 18 teams). The task assigns each of 27,972 hidden-test Arabic segments one of 6 broad and one of 74 specific genres, given only as English prose definitions, with the specific-label inventory disjoint from training. Our pipeline updates no model parameters but is not free of task-specific effort: the judge prompt carries 74 hand-written cues, and components were selected on the public leaderboard. A family-restricted, definition-guided LLM judge scores candidate genres setwise under a hierarchically constrained decoding scheme, whose five-step pipeline reaches 0.7139. That base led on Specific Macro F1 but trailed on Broad Macro F1, so we add a correction stage: two instruction models re-predict the specific genre, each shown the entire 74-definition taxonomy, and a label is edited only where both agree against the base. That system scores 0.892 Broad Macro F1, alongside fewer topic-trap calls, where the subject a text discusses displaces the genre it instantiates. In a post-evaluation audit of instances that the base labelled as book descriptions, Religious calls fall 394 → 233 → 119 as the prompt adds the type-versus-topic convention and then the full taxonomy. Refinements and an attractor drain reach 0.7352.

Harmful content appears in diverse forms, ranging from multimodal memes targeting protected groups to textual prompts designed to elicit unsafe responses from large language models. However, existing Arabic resources typically study these settings separately and often rely on coarse-grained labels. We introduce **ArGuard**, a shared task on harmful content detection in Arabic memes and LLM prompts. ArGuard consists of two tracks: Track A focuses on multimodal hate detection in Arabic memes, while Track B addresses harmful prompt detection for Arabic LLM safety evaluation. Overall, 58 teams registered for the shared task, 35 participated in the final evaluation phase, and 27 submitted system-description papers. Participating systems explored a range of approaches, including fine-tuning encoder- and decoder-based models such as AraBERT, Jais, and Qwen3-VL. The best-performing systems achieved macro-F1 scores of 0.823 on Subtask A1, 0.419 on Subtask A2, 0.984 on Subtask B1, and 0.790 on Subtask B2. These results highlight fine-grained meme classification (Subtask A2) as the most challenging setting, partly due to sparse labels and train-test distribution shifts. To facilitate further research on Arabic harmful-content detection, we release the task datasets and associated resources through the shared-task website.

This paper describes the participation of the dbilianos team in the ArGuard 2026 Shared Task on harmful prompt detection for Arabic LLM safety evaluation. We participated in both subtasks: binary safe/unsafe classification (Task B1) and granular harm category classification (Task B2). Our approach systematically tested a central hypothesis: whether translating Arabic prompts into English improves classification performance. We experimented with multiple pipelines, including classical machine learning with TF-IDF features, zero-shot and few-shot prompting with DeepSeek-V4, and hybrid approaches. Our results reveal a divergence across subtasks. For binary classification, DeepSeek zero-shot on English translations achieved 0.89 macro-F1, securing 4th place in this subtask, and confirming that LLM-based prompting is highly effective for coarse-grained safety decisions. For granular classification, however, the best performance came from training a Support Vector Machine directly on original Arabic text using TF-IDF features, which achieved 0.61 macro-F1, placing the team 16th in this subtask. This outperformed our translation-based strategies, suggesting that fine-grained harm categorization is highly sensitive to linguistic and cultural nuance that is lost or distorted during translation. The translation hypothesis was only partially supported locally, and ultimately reversed on the final test set for both subtasks. We discuss implications for cross-lingual safety evaluation and highlight the limitations of translation-based pipelines for tasks requiring subtle intent interpretation.

Arabic harmful-prompt detection has remained challenging due to dialectal variation, implicit expressions and context-dependent intent. In the ArGuard 2026 Shared Task, we have participated in Subtask B1, a binary safe/unsafe classification task scored by macro-F1. We have identified evaluation leakage in an initial pipeline: the development set has been reused for early stopping, threshold tuning and ensemble optimization, which has inflated development scores while masking poor generalization on the blind test set. To correct this, we have introduced a leakage-controlled protocol with separate fitting, checkpoint-selection and tuning splits and have compared a TF-IDF SVM baseline, MARBERTv2 and a stacking ensemble, selecting the best-performing method based on tuning-set results. We have further applied five rounds of error-driven data augmentation to address recurring failure patterns. Our final system has achieved 0.861 accuracy, 0.878 precision, 0.862 recall and 0.860 macro-F1 on the blind test leaderboard, ranking 5th. Manual error analysis has shown that remaining errors have mainly stemmed from false positives on sensitive but non-harmful content, confirming that leakage-controlled evaluation has been necessary for a reliable blind-test estimate.

We present the L3IA Morocco team’s system for shared Task B, designed for large language models and focused on detecting harmful prompts in Arabic. In this context, we define two subtasks: B1 and B2, which refer, respectively, to the binary safe/unsafe classification and the classification of harmfulness into 7 categories. Our system combines three fine-tuned Arabic encoders: AraBERTv2, CAMeLBERTmix, and MARBERT, using soft voting that takes into account the predominant classes. It achieves a macro-F1 score of 0.946 for B1 and 0.902 for B2 during the development phase. In contrast, performance drops to 0.639 for B1 and 0.758 for B2 in the blind test. We demonstrate that this generalization divergence is due to a "unsafe” overprediction bias resulting from an unbalanced training corpus (80% unsafe) and not due to alignment of identifiers or dialects. Finally, we observed a negative result: in zeroshot and few-shot, AceGPT-7B, a generative Arabic LLM, performed poorly compared to fine-tuned encoders (0.38-0.42 on Darija).

We present AK_QMUL’s submissions to the ArGuard-2026 shared tasks on Arabic harmful-prompt detection: a binary safe/unsafe task (B1) and a seven-way harm-domain task (B2). Both submissions use one backbone – JAIS-6.7B fine-tuned with 4-bit QLoRA and a linear classification head over its last-token representation. For B1 we train a four-way head that factors the label by dialect register safe, unsafe × MSA, other, we augment the training fold with MSA prompts from AraSafe, and decode with a plain argmax to score safe/unsafe. For B2 we fine-tune the same backbone as a seven-way classifier and average the softmax outputs of five random seeds. Both submissions ranked first on their track, with macro-F1 0.984 on B1 and 0.790 on B2. We report results by dialect and by class: on B1 the model is accurate across all six dialects (93.6–99.8%) and its predicted unsafe rate tracks the true rate closely; on B2 six of the seven classes reach F1 between 0.73 and 0.98, with the residual error concentrated in the Bully class, which overlaps semantically with harassment and harm-to-others.

The rapid growth of social media and generative AI has significantly increased the volume of Arabic textual content, memes, and user prompts shared across online platforms. While these media facilitate communication, creativity, and information exchange, they can also be exploited to disseminate hate speech, harmful content, and unsafe instructions. As a result, there is an increasing need for reliable automated methods to identify hateful Arabic memes and harmful (unsafe) prompts. This paper leverages pre-trained and fine-tuned multilingual embedding models to learn rich semantic representations from textual and visual content. The evaluation is conducted using a dataset of Arabic memes and prompts proposed in the ArabicNLP ArGuard 2026 challenge. The results underscore the capacity of large language model (LLM)-based embedding models to deliver superior performance. Developing robust detection systems is essential for improving content moderation and fostering a more secure and trustworthy online environment for Arabic-speaking users.

We describe our system for the ArGuard 2026 shared task (Task B), harmful-prompt detection for Arabic, with subtasks B1 (binary safe versus unsafe) and B2 (seven-way harm-domain classification), both scored by macro-F1. Our system pairs a Unicode-normalization front-end with an ensemble of fine-tuned Arabic encoders, treats B2 as topic-domain classification over all prompts, and adds a retrieval-augmented resolver for the confusable interpersonal classes; it reaches development macro-F1 0.907 (B2) and 0.950 (B1), and on the blind test B1 0.9839 (1st, tied) and B2 0.7875 (2nd). Our main contribution explains the large B2 development-to-test gap. One fifth of the B2 test is a "novel" slice: AraSafe prompts re-annotated into the seven ArGuard domains under a fixed cross-taxonomy mapping that the organizers confirmed. Because that mapping does not preserve the training convention and is recorded in no data available during the task, no system can infer it, and the competitive B2 field compresses onto a blind 0.68-0.79 band, far below the ~0.90 reachable in-distribution. On the familiar slice we score macro-F1 0.908, matching development; we argue this is the honest measure of the system.

For the ArGuard 2026 shared task on harmful content detection in Arabic memes (Task A) and LLM prompts (Task B), we fine-tune Arabic transformer encoders across all four subtasks. Memes additionally use a vision tower (ViT-B/16 for the binary subtask, CLIP ViT-L/14 for the fine-grained one) late-fused with MARBERTv2; ensembling choices and decision thresholds are tuned per subtask throughout. Two methodological choices account for much of our gains: selecting checkpoints by development macro-F1 rather than loss, and aligning the B2 training target with the evaluation distribution. Our largest single effect, however, was diagnostic. The B1 blind test set is approximately class-balanced while its training data is ~80/20, so our dev-tuned threshold collapsed macro-F1 from 0.9557 on dev to 0.6345 on test. We recovered the test class balance algebraically from the aggregate metrics the platform reported to all participants; re-thresholding the same saved probabilities, with no retraining, restored 0.8628. Because this uses only aggregate leaderboard feedback (within the task rules), the B1 figure is not directly comparable to teams that did not use the channel and is not a generalisation estimate. This correction is also not general: on the weaker multi-label classifier (A2), stronger prior matching reduced performance in every setting we tested, so matching the test prior is not necessarily F1-optimal. We additionally report a catalogue of quantified negative results, including external-data augmentation with AraSafe. Final test macro-F1: A1 0.7869, A2 0.3463, B1 0.8628, B2 0.7622.

This paper describes CalibGuard, our system for the ArGuard 2026 Shared Task on harmful content detection in Arabic memes and large language model prompts. We participate in all four subtasks: binary and fine-grained hateful meme classification (A1, A2) and binary and harm-domain prompt classification (B1, B2). For Task A, we extract frozen embeddings from three vision-text encoder pairs and train a joint two-headed classifier under 5-fold cross-validation with seed averaging. For Task B, we fine-tune AraBERTv02 with class-weighted cross-entropy. CalibGuard obtains 75.53% macro-F1 on subtask A1, 36.53% on A2, 65.38% on B1 and 68.85% on B2, ranking 11th, 3rd, 15th and 13th respectively. We further show that the B1 test split contains 49.3% unsafe prompts against 79.8% in training, and that this distribution shift accounts for a 29-point gap between our development and test scores, concentrated entirely in Modern Standard Arabic. Rethresholding to the test prior recovers 13.7 points without retraining. We additionally report two negative results: constrained decoding marginally reduces A2 performance, and restricting B2 training to unsafe prompts costs 5.7 points.

We present SZED’s submission to ArGuard 2026 Track A: Arabic hateful meme detection. The task has two parts: binary classification (A1) and fine-grained multi-label tagging (A2). For A1, we fuse MARBERTv2 and CLIP ViT-L/14 embeddings, comparing concatenation, self-attention, and ensemble heads. For A2, we use a multi-task label-attention architecture with per-class query vectors and an auxiliary head for cascade calibration. We reach macro-F1 = 0.7597 on A1 and 0.3142 on A2. We report reproducibility across runs, per-class error patterns, and the effect of cascade calibration on the long-tailed A2 labels.

The classification of Arabic prompts as Safe or Unsafe is the focus of this submission to ArGuard 2026 Subtask B1. We developed a lightweight hybrid classifier that combines lexical and semantic information. The lexical branch uses normalized word-level and character-level features with a class-balanced LinearSVC classifier. The semantic branch uses frozen multilingual E5 embeddings with a class-balanced logistic regression classifier. We combined the outputs of both branches using late fusion. On the development set, LinearSVC achieved a Macro-F1 score of 0.93, while the frozen E5 branch achieved 0.88. The final hybrid system achieved a Macro-F1 score of 0.93. On the test set, the submitted system achieved 0.78 accuracy, 0.83 precision, 0.70 recall, and 0.72 Macro-F1.

Detecting hate speech in social media memes requires disentangling complex cross-modal interactions, a challenge magnified in low-resource and morphologically rich languages like Arabic. We present Team Axiom’s submission to the ArGuard 2026 Shared Task (Task A1) on binary hate meme classification. While conventional multimodal wisdom posits that text and visual features must be jointly fused, empirical evidence across non-English domains reveals that noisy visual backbones and cross-modal space misalignment frequently degrade language signals. We validate this hypothesis by fine-tuning a unimodal MARBERTv2 sequence classifier without external multimodal fusion. Our simple, well-regularized approach achieved a Macro-F1 of 0.770 and an accuracy of 0.818 on the official gold test set, securing 6th place on the global leaderboard and outperforming both the official AraBERTv2 baseline and our own complex multimodal fusion architectures.

The classification task for ArGuard 2026 Track A, Subtask A1, identifies whether an Arabic meme is Hateful or Not Hateful. Our system combines three separately trained models. In the first experiment, we use word-level and character-level features with balanced linear classifiers. In the second experiment, we fine- tune MARBERT for text classification using class-balanced focal loss. In the third experi- ment, we use a frozen SigLIP image encoder with balanced logistic regression. We combine the Class-1 probabilities from the three mod- els using late fusion. On the test set, the fused model achieved a Macro-F1 score of 0.76 and an accuracy of 0.79

This paper presents SATLab’s participation in the ArGuard 2026 shared sub-task on identifying seven types of harmful LLM prompts in Arabic. The proposed system is language-agnostic and requires no resources other than the training data provided in the shared task. It is based on character n-grams and a logistic regression classifier. Performance on the Test set (Macro-F1 = 0.736) was average, approximately 0.05 lower than that of the top teams (0.790) and approximately 0.05 higher than the baseline (0.687). Notably, this performance was substantially lower than those achieved in cross-validation and on the Development set. Several analyses aimed at understanding these discrepancies are reported. An analysis of the confusion matrix indicates that the “Harm-to-other” category poses the greatest problem, likely because it is partially compatible with all other categories except “Adult-content”.

We describe the QuadCore submission to Sub task B2 of the ArGuard 2026 shared task, which asks systems to assign an Arabic prompt addressed to a large language model to one of seven harm domains. Our system is a soft-voting ensemble of three Arabic en coders with complementary pre-training cor pora- MARBERTv2, CAMeLBERT-DA and AraBERTv02-Twitter- each fine-tuned under two random seeds, with the annotated dialect prepended to the prompt as a textual condition ing prefix; it uses no external data, no augmen tation and no generative models. It reached 0.7665 macro-F1 on the official test set, rank ing 4th of the 16 teams that submitted to the subtask. Using the gold labels released after the evaluation period, we then show that this sin gle number conflates two regimes: on the first 4,809 test items, whose label prior, dialect mix and lexical profile match the training data, the same predictions reach 0.9044 macro-F1, while on a contiguous final block of 1,170 items they collapse to 0.2073. That block is entirely Mod ern Standard Arabic, contains no adult-content instance, and its nearest-neighbour label agree ment with the training set is 0.1829 against 0.7777 elsewhere- close to the 1/7 chance rate. A substantial portion of our system’s errors on the official test set is therefore attributable to this shift rather than to the difficulty of harm domain classification itself.

We present our approach in this submission for ArGuard 2026 Subtask B1, which classifies Arabic prompts directed at large language models as Safe or Unsafe. Our system combines four sparse TF-IDF feature representations with dialect indicators and surface-level features. We trained logistic regression and Naive-Bayes SVM (NBSVM) classifiers and combined their predictions using late fusion. The system achieved a Macro-F1 score of 0.94 and an accuracy of 0.96 on the development set. On the test set, our submission achieved 0.79 accuracy, 0.85 macro precision, 0.71 macro recall, and 0.72 Macro-F1.

Hateful content in memes is difficult to detect automatically, as harmful intent often emerges from the interaction between image and text rather than either modality alone, a challenge compounded in Arabic by dialectal variation and sarcasm. This paper presents the system submitted by team Bahash-AI to Subtask A1 (binary hateful meme detection) of the ArGuard shared task at ArabicNLP 2026. A late-fusion classifier combining MARBERTv2 and ViT-base is first developed and evaluated as a standalone development-phase system. Building on this, the final submission ensembles a CLIP-based text-image classifier with this same late-fusion classifier, retrained on the combined train, development, and development-test partitions. Class imbalance (approximately 30% Hateful) is addressed via class-weighted losses, oversampling, and threshold selection. The final ensemble achieves a Macro-F1 of 0.7457 and an accuracy of 0.7680 on the blind test set, ranking 12th of 16 teams on the official leaderboard and outperforming the development-phase system by 9 Macro-F1 points. Error analysis using per-class precision and recall shows that the two systems reach a similar Hateful-class F1 gap through opposite error patterns.

This paper presents the CIC system for ArGuard 2026 Subtask B2, which classifies unsafe Arabic prompts into one of seven harm domains. Our system combines four complementary components: a fine-tuned CAMeLBERT-Mix model, a modified-Huber stochastic gradient descent classifier, a character-level nearest-neighbor model, and a character class-centroid similarity model. On the development set, we achieved a Macro-F1 score of 0.916 and an accuracy of 0.923. On the test set, the proposed model achieved an accuracy of 0.766, a macro precision of 0.760, a macro recall of 0.763, and a Macro-F1 of 0.751. The results show that combining contextual transformer representations, lexical features, and retrieval-based similarity signals can improve fine-grained harmful-prompt classification.

We describe our submissions to Task A of the ArGuard 2026 shared task on harmful-content detection in Arabic memes: binary hateful-meme detection (Subtask A1) and fine-grained multilabel categorization (Subtask A2). Our A1 system is a probability-averaging ensemble of 22 fine-tuned models: 18 Qwen3-VL vision-language members (8B and 32B) trained with LoRA and scored via first-token logits, and 4 Arabic text-encoder members (MARBERTv2, AraBERTv02-twitter), combined at a fixed 0.85/0.15 modality mix. The system ranked first on Subtask A1 with 0.8234 macro F1, 2.2 points above our best single member. For A2, six LoRA-tuned VLMs with sigmoid multilabel heads, an A1-consistency mask, and flat 0.5 thresholds scored 0.3900 macro F1. Post-hoc analysis on the released gold labels shows that our single global threshold survived an unannounced 8-point drop in the share of hateful memes in the test set, and that most remaining errors in both subtasks separate benign sarcasm or humor from mockery aimed at a protected target.

Detecting hateful content in Arabic memes is challenging due to dialectal variation, sarcasm, cultural references, and the interplay between visual and textual signals. We participate in ArGuard 2026 Track A (Multimodal Hateful Meme Detection) with a multimodal architecture that fuses AraBERT textual representations, ResNet50 visual features, and a lightweight knowledge-graph (KG) ontology vector derived from domain-specific Arabicbinary keywords. The model jointly optimizes (Hateful / Not Hateful) and multi-label finegrained classification through a gated fusion layer with strong regularization (dropout 0.45, weight decay 0.05, label smoothing 0.1). Gradual unfreezing of AraBERT layers, weighted sampling, and Arabic-specific normalization further improve robustness. On the official development set the system reaches a binary macro-F1 of 0.710; on the blind test set it obtains binary macro-F1 0.717. We analyse the contribution of each modality and discuss remaining challenges of fine-grained multi-label prediction under severe class imbalance.

Detecting harmful prompts directed at large language models is a critical safety challenge, especially in Arabic where dialectal variation, code-switching, and culturally specific expressions of harm complicate classification. The ArGuard 2026 shared task addresses this problem through two subtasks: binary classification of prompts as safe or unsafe (B1) and fine-grained classification of unsafe prompts into seven harm domains (B2). We participate with a multi-task fine-tuning of MARBERT that shares encoder representations across a binary head and a seven-class granular head. Training combines focal loss with class weights and label smoothing, a hierarchical consistency regularizer that discourages contradictory safe/unsafe and granular predictions, a weighted sampler, and light character-level augmentation. On the official development set the system reaches a binary macro-F1 of 0.9618 and a granular macro-F1 of 1.0000; on the blind test set it obtains binary macro-F1 0.688 and granular macro-F1 0.752. We analyse the substantial development–test gap and discuss limitations of the submitted system.

This paper presents the systems we developed for ArGuard 2026 Task B, addressing Arabic prompt safety through binary safe/unsafe classification (B1) and fine-grained assignment of unsafe prompts to harm domains (B2). Our central strategy is cross-lingual data augmentation: we machine-translate a large English prompt-safety corpus into Arabic and reuse it across both subtasks. For B1 we fine-tune the Arabic-native ALLaM-7B on self-generated rationales using the granular labels from B2 as signal. We compare against label-only training; our submission ranks 2nd (macro F₁=0.924). The experiments show the augmentation drives accuracy and not the self-generated chain-of-thought. For B2 we fine-tune Arabic encoders (full and adapter tuning) with a two-stage "warm’ pipeline that first trains the encoder on the easier binary task, before specialising it to the seven domains; our submission ranks 5th (macro F₁=0.762)

We describe our systems for all four ArGuard 2026 subtasks: binary and fine-grained Arabic hateful-meme detection (A1/A2) and binary and granular harmful-prompt detection (B1/B2). Our systems are deliberately conventional (a LoRA-tuned Qwen3-VL-8B for memes, a nine-encoder ensemble for prompts) because our central finding is that most of the achievable headroom lay not in the model but in decode-time calibration: per-class thresholds refit under structural constraints, rank-based decoding against known or published class priors, and threshold transport under distribution shift. On subtask A2, our submissions rose from 0.248 to 0.419 macro-F1; roughly 0.15 of that gain came from decode-time changes against under 0.03 from model-side interventions, and a fixed-model ablation on the ensemble’s char-ngram member attributes +0.070 to the decode stack alone. On B1, recalibrating a single threshold recovered +0.16 macro-F1 after we detected that the blind test was rebalanced to approximately 49% unsafe against an 80% unsafe training prior. Our final systems ranked 1st on subtask A2 (macro-F1 0.4194; 0.3913 without per-class scoring feedback), 3rd on A1 (0.7889), 3rd on B2 (0.7695), and 7th on B1 (0.7552). We additionally report a benchmark audit conducted during participation: a sequential-id label leak in the Task B development release (macro-F1 1.000 from ids alone), identified before the test release, left unexploited, and absent from the re-indexed test files; latent taxonomy structure that contradicts the task description; and a quantified gap between development and test distributions.

Multimodal hate speech detection requires interpreting how an image and its accompanying text jointly construct a target, stance, or derogatory implication. In Arabic memes, this relation is often shaped by locally recognizable public figures, political references, visual analogies, and indirect humor, making either modality insufficient on its own. We tackle this problem in Track A of ArGuard 2026, covering binary hateful meme detection (A1) and fine-grained multi-label classification (A2). Our experiments compare jointly trained dual encoders, parameter-efficient vision–language model (VLM) adaptation, and training-free prompting. We fine-tune Qwen3-VL, Llama 3.2 Vision, and Gemma 4 with QLoRA using task-specific binary and multi-label objectives. We also build CLIP–MARBERTv2 and SigLIP–AraBERT-Twitter cross-attention systems and evaluate frozen VLMs with zero-shot prompting and in-context learning. Qwen3-VL achieves our highest A1 score (0.7610), while SigLIP–AraBERT-Twitter leads A2 (0.3851), where label sparsity and multi-label prediction remain challenging.

We present our systems for ArGuard 2026 Track A, which aims to detect harmful content in Arabic memes. The track includes two tasks: a binary task for identifying hateful memes (A1) and a multi-label task for predicting specific categories (A2). For A1, our approach combines a fine-tuned Qwen3-VL sequence classification model, Arabic text encoders, and a Qwen text LoRA classifier. We then apply top-k calibration and make targeted adjustments based on model disagreements. On the official test set, our system reached a macro-F1 score of 0.7896 and an accuracy of 0.8200. For A2, we found that using thresholds from the development data caused us to overpredict common labels and miss rare ones. Our best A2 submission uses an ensemble of six Arabic text models, with label-frequency-aware top-k calibration and rare-label boosting. This approach achieved a macro-F1 of 0.3601 and a micro-F1 of 0.5150. Overall, we found that careful calibration of a few fine-tuned Arabic and multimodal classifiers contributed more to our results than using larger zero-shot vision-language models.

We describe our system for Subtask A1 (binary hateful meme detection) of the ArGuard 2026 shared task (Alam et al., 2026). Our final system fine-tunes MARBERTv2 with class-weighted loss and macro-F1-based checkpoint selection, then applies a post-hoc decision-threshold calibration to correct for a class-distribution shift we found between the training/development splits (37.8% Hateful) and the held-out test split (29.6% Hateful). This calibration step requires no retraining and adds 3.3 macro-F1 points (0.7286 → 0.7616) on the official test split. On the same test split, this calibrated system also outperforms two alternatives we evaluated there – multi-seed ensembling and naive multimodal late-fusion. They match or exceed backbone-substitution and multi-model-ensemble variants that we were only able to evaluate on local development data. We report our full experimental log, including the negative results, with each result explicitly labeled by evaluation split, and compare our score to the AHA-Memes benchmark, where a much larger vision-language model reaches a similar range. Code and configuration files are released for reproducibility.

StanceEval 2026 is the second edition of the StanceEval shared task series on stance detection in Arabic social media text. Stance detection aims to identify a writer’s stance toward a given topic. Given a tweet and a target, participating systems must determine whether the writer’s stance is Favor, Against, or None. This edition focuses on cross-target generalization across two distinct evaluation tracks: Track 1 evaluates thematically related cross-target transfer (testing on Women Driving, related to Women Empowerment in the training data), while Track 2 evaluates cross-domain transfer to completely unseen targets (E-Cars and Trimester System). The shared task attracted 80 registered teams from 12 countries. During the evaluation phase, 30 unique teams submitted entries, with 21 teams officially ranked in Track 1 and 13 in Track 2 following validation filtering, and 20 teams submitting system-description papers. Participating teams employed diverse methodologies, including fine-tuned pretrained language models, prompt-based and retrieval-augmented large language models (LLMs), fine-tuned LLMs, and hybrid cascades. Top systems achieved impressive F_avg2 scores of 0.8994 on Track1 and 0.9400 on Track2, substantially outperforming the strongest baselines (0.7366 and 0.7475, respectively), where F_avg2 denotes the macro-averaged F1 score over the Favor and Against classes. Counterintuitively, performance on the unseen targets was higher than on the related target, a disparity could be driven by extreme target polarization, class imbalance, and dialectal or sarcastic nuance across topics.

Arabic-language stance detection remains challenging, and previous shared-task systems have largely relied on multitask learning and ensembles. While these systems achieve state-of-the-art performance, their applicability and transferability are limited by the additional complexity introduced by multitask learning. To reduce this complexity, we introduce CLASP-Ar, which reformulates the task as cloze-style masked language modeling. In this approach, the target, predicted sentiment, and text are combined into a single prompt whose [MASK] prediction is restricted to a verbalizer-constrained label vocabulary. When evaluated on the StanceEval-2026: Arabic Stance Detection Shared Task, our proposed approach achieves an F_avg2 score of 71.36% on Track 1 and 74.14% on Track 2 test sets.

We describe our systems for both StanceEval- 2026 Arabic stance-detection tracks. Our MARBERT-centered pipeline optimized the primary macro-F1 over AGAINST and FAVOR directly, combining target-aware fine-tuning, probability ensembles, and guarded transfer experiments. We report one official test sub- mission per track. The Track 1 target-routed MARBERT system scored 0.8925 on devel- opment and 0.6743 on the official test set. The Track 2 MARBERT/Fanar/Qwen3 system scored 0.8360 on development and 0.8100 on the official test set. The released Track 1 evaluation input contained only Women Driv- ing, absent from its labeled train and develop- ment targets, making the nominal seen-target track an effective target-transfer test for our system. These results show that validation pro- tocols must reflect the intended target-transfer setting; otherwise, same-target development scores can substantially overstate transfer per- formance.

Cross-target Arabic stance detection presents a unique challenge due to the rich morphological complexity, regional dialects, and heavy reliance on sarcasm within social media discourse. In this paper, we describe the system submitted by Team Axiom for Track 1 (Seen Targets) of the StanceEval-2026 shared task. Our primary contribution is a resource-efficient, Two-Stage Hybrid Cascade architecture. This system utilizes a fine-tuned MARBERTv2 encoder as a high-confidence gatekeeper, dynamically routing uncertain and complex tweets to a locally hosted, 4-bit quantized Large Language Model (Llama-3.1-8B-Instruct). To prevent label bias and ground the LLM’s reasoning, the fallback stage incorporates a dynamically balanced, class-specific TF-IDF sub-word retriever. Our official submission achieved an F_avg2 score of 0.6910 (Rank 18), demonstrating a measurable improvement over our standalone encoder baseline (0.6722).

We present Tarjih, our proposed system for the Stance Detection in Arabic Language shared task (StanceEval-2026). Rather than predicting stance directly, Tarjih formulates the task as verification among competing arguments. An instruction-tuned LLM generates one steelman argument for each stance label, and a second blind pass evaluates each argument against the tweet without access to the gold label. Each argument–tweet is then jointly encoded with AraBERT to model their compatibility. From this pipeline, we derive three systems: a gated classifier, a pair scorer, and a ranking read-out. The pair scorer achieves the best seen-target test Favg2 of 0.73, while the gated classifier performs best on the unseen-target track with 0.77. Overall, the lower test performance and ranking among the submitted systems indicate limited generalization across targets.

This paper presents a system for StanceEval-2026 Track 1, Arabic stance detection on seen targets. Systems in this track are trained on the provided targets and evaluated on a related target within the same thematic domain. Open large language models are fine-tuned with QLoRA on the original Mawqif training data and adapted to the test target through transductive self-training, in which unlabeled test tweets are added to training only when independent models agree on their label. The final prediction is a majority vote over three fine-tuned models, Qwen2.5-7B, the Arabic-native ALLaM-7B, and a sarcasm-aware variant that predicts the official sarcasm label before the stance label. The system scored 0.7564 Favg2 and placed 12th of 21 teams on the official leaderboard. Across seventeen scored variants, task fine-tuning outperformed zero-shot prompting at every model scale, heterogeneous fine-tuned models converged on 98% of test predictions, and further gains required decorrelated ensemble members. The sarcasm annotation supplied this decorrelation and improved the vote, while the sentiment annotation degraded it.

Our two systems for StanceEval-2026, Arabic stance detection on Mawqif-XT, placed 1st of 31 on Track 1 (Favg2 = 0.8994) and 2nd of 22 on Track 2 (0.9357). Track 1 averages six zero shot readings of each tweet by frontier models from two providers, and calls a fine-tuned Arabic pipeline only to break exact ties. Track 2 is one reader, given a short generated description of the target so it can score stance expressed through a related topic. The leaderboard publishes three aggregate scores over the same predictions, and reading them shows that about 60% of our errors on both tracks are tweets annotated NONE that we placed in a stance class. Because the score ignores that class, the gain from moving such a row is fixed before the row is read. We derive the precision a repair must reach, and nine rankings plus an oracle from the corpus’s own annotation columns all fall short of it.

This paper presents the SRK AI Lab system for Arabic stance detection in the StanceEval 2026 shared task using the Mawqif-XT benchmark. We employ a few-shot multi-LLM framework with GPT-5.5 Instant, Gemini 3.6 Flash with Extended Thinking, and Qwen 3.7 Plus to classify target–post pairs as Favor, Against, or None. The models independently predict stance, followed by GPT-5.5 Instant auditing based on target–post relation, linguistic context, and stance evidence. The audit results are combined with cross-model comparison to select the final label. Our system achieves an official Favg2 of 0.8420 in Track 1 and 0.9192 in Track 2, ranking 8th and 3rd, respectively. These results demonstrate the performance of multi-LLM few-shot prompting with GPT-5.5 auditing for Arabic stance detection.

This paper describes the system submitted by the phonghoang team for Track 1 (Seen Tar- gets) of the StanceEval-2026 shared task on Arabic stance detection, where models are trained on provided targets and evaluated on a related target within the same thematic do- main. We propose a two-stage framework to address this within-domain target shift. In the first stage, we fine-tune a Multi-Task MAR- BERT model that jointly learns stance, senti- ment, and sarcasm, further regularized with R- Drop and Stochastic Weight Averaging (SWA), achieving Favg2 of 0.8561 on the development set. In the second stage, we adapt to the test target (“Women Driving”) by combining con- trastive few-shot pseudo-labeling via a Large Language Model, fusion with an external Ara- bic stance dataset, and high-confidence samples filtered from a thematically related training tar- get. The resulting augmented training set is used to adapt our optimal Stage 1 architecture, achieving an official Favg2 of 0.8431 on the test set.

Stance detection identifies an author’s position toward a specified target, yet Arabic social-media text remains challenging because stance is often implicit and shaped by dialect, sarcasm, informal expression, and culturally specific references. StanceEval-2026 evaluates this problem in two transfer settings: a related target within a seen thematic domain (Track 1) and entirely unseen topics (Track 2). We compare four modeling strategies: single-task Arabic encoder fine-tuning, multi-task learning with sentiment and sarcasm supervision, 4-bit QLoRA adaptation of the ALLaM-7B model, and Retrieval-Augmented In-Context Learning (RAG-ICL) with frozen large language models (LLMs). Our strongest single-backend configuration is a data-centric RAG-ICL pipeline combining label-balanced retrieval, hashtag and negation adaptation, text cleaning, similarity filtering, and target descriptions. A majority-vote ensemble of this configuration across three LLM backends obtains official F_avg2 scores of 0.8719 on Track 1 and 0.9016 on Track 2, ranking 3rd and 7th, respectively. Under the shared-task test conditions, the retrieval-based systems substantially outperform the supervised systems evaluated here, particularly on Track 2.

Arabic stance depends on quotation, sarcasm, negation, hashtags, and target scope, which surface polarity may miss. For StanceEval-2026, we study Codex and Gemini prompting through two submission systems: a replicated contextual Codex predictor and a gated judge that retrieves labeled examples using seven AI-generated discourse cues. On 619 Dev examples, cue retrieval changes Favg2 from 0.8580 to 0.8613 for GPT-5.6-luna, 0.8767 to 0.8844 for Gemini 3.5 Flash, and 0.8646 to 0.8554 for Flash-Lite. Random and semantic controls show no consistent retrieval advantage. On the 352-row hidden set, the best retained Codex member scores 0.8685 Favg2 and currently ranks fourth, while gated retrieval scores 0.8604. Codex and OPUS-MT translation pipelines score 0.7848 and 0.7304, and a synthetic-data classifier scores 0.7334. These results support replication and retrieval as auditable system components, not as general performance gains.

We present our system for StanceEval-2026 Track 1, which focuses on Arabic stance detection toward Women Driving using the Mawqif-v2 dataset with three labels: Favor, Against, and None. We compare fine-tuned AraBERT models with zero-shot Gemma-4-31B. We refine the Gemma prompt using recurring development-set errors, focusing on dialect and indirect meaning, target identification, sentiment versus stance, hashtags, sarcasm, and the distinction between reporting and endorsement. We also compare the calibrated system with several alternative inference strategies, including self-consistency, model debate, and dialect-interpreter pipelines. The final Gemma-4-31B system achieves 86.61% Favg2, 67.46% Favg3, and 82.67% accuracy on the official blind test, ranking 5th. Our results show that target-specific context and clear decision rules can be more effective than adding additional inference stages or model interactions. The main remaining challenge is distinguishing Against from None, particularly when stance is indirect, ambiguous, or expressed through humor.

We present Nabra, a system submitted to StanceEval-2026 Track 1 for Arabic stance detection. Our approach fine-tunes MARBERTv2 on the Mawqif-v2 dataset, exploring seven configurations including class weighting, regularization, data augmentation via back-translation, resampling strategies, and zero-shot inference with Qwen2.5-7B. Our best model by the primary metric Favg2 is M1 (Baseline), achieving 0.837 on the development set and 0.6704 on the official blind test set. Error analysis reveals that Favor→Against misclassification accounts for 84.6% of errors, driven by negative sentiment words co-occurring with supportive stances. We find that metric choice (Favg2 vs. Favg3) significantly impacts model selection decisions.

We present CaRPS, our system for Track 1 of StanceEval-2026. CaRPS combines a target-conditioned encoder with an LLM-based labeling and verification process to handle targets that are not seen during encoder fine-tuning. The encoder takes both the target and the tweet as input, while the LLM is used to generate labels and verify difficult predictions. These LLM-generated labels can then be used as pseudo-labeled data to further fine-tune the encoder, with the goal of gradually reducing the need for LLM calls. On the development leaderboard, our target-conditioned encoder alone achieves 87.19% Favg2, 76.56% Favg3, and 85.30% accuracy. For the official evaluation, we used adversarial LLM prompting for the target not seen during encoder fine-tuning, achieving 82.89% Favg2, 62.67% Favg3, and 78.98% accuracy. Additional experiments show that adversarial verification improves stance and sarcasm classification over one-shot prompting, while pseudo-label self-training helps adapt the encoder to a held-out target. Overall, CaRPS explores a practical combination of efficient encoder-based prediction and LLM-based labeling and verification for Arabic stance detection.

This paper describes our system for StanceEval 2026 Track 2, which requires predicting stance toward targets not seen during training. We explored fine-tuning different Arabic BERT encoder models, using class weighting techniques, incorporating auxiliary sentiment and sarcasm tasks in sequential and parallel multi-task setups, and focal and margin-based contrastive losses. We also tried zero-shot and few-shot prompting, self-consistency with escalation, LoRA fine-tuning, and an ensemble approach. Our best-performing submission on the official blind test set was a training-free few-shot prompting approach using Gemini 2.5 Flash with an extended prompt, achieving an F_avg2 of 87.41 and an F_avg3 of 76.94, and ranking 8th out of 13 teams. This result shows that a simpler, training-free prompting approach can generalize to unseen targets more effectively than fine-tuned and ensemble systems.

We describe our entry to both tracks of StanceEval-2026, an Arabic stance-detection task built on the Mawqif-XT corpus. Track 1 (Seen Targets) is evaluated on Women Driving, while Track 2 (Unseen Targets) is evaluated on Electric Cars and the Trimester school calendar system. Rather than committing to a single approach, we compare two paradigms: fine-tuned Arabic encoders trained with multi-task learning using Mawqif-XT’s sentiment, sarcasm, and rationale annotations, and native Arabic LLMs used in a few-shot setting without any fine-tuning. For each track, we build the final system around the stronger approach. Across more than forty experimental configurations, few-shot LLMs consistently outperform fine-tuned encoders on both tracks. Our final systems achieve an Favg2 of 0.834 on Track 2 (ranked 9th) and 0.747 on Track 1 (ranked 13th), the latter through a complementary fusion of ALLaM-7B and Fanar-9B. We further observe that model performance may be related to how well pretraining coverage aligns with a target’s cultural and linguistic domain, that fusing two comparably strong models can outperform either model individually, and, perhaps most importantly, that more elaborate methods never surpass a strong LLM with a simple prompt. We release our code and prompts to support reproducibility.

Arabic stance detection is a task that identifies whether an Arabic text expresses Favor, Against, or None toward a given target. It is important for understanding public opinion and social media discourse, particularly in Arabic where dialectal variation, sarcasm, and informal expressions make stance interpretation difficult. To address this, we took part in StanceEval-2026 Track 1, where models are trained on provided targets and evaluated on a related target from the same thematic domain. We investigate two complementary approaches: fine-tuning-based systems using Arabic encoders and parameter-efficient LLM adaptation, and prompted GPT-4o without fine-tuning. Our experiments include encoder ensembles, multi-task and contrastive learning, attention aggregation, augmentation, calibration, QLoRA, and prompted LLMs with few-shot demonstrations, target-specific context, structured reasoning, and self-consistency voting. The best fine-tuning-based system achieves 0.701 Favg2, while GPT-4o v4 reaches 0.827 Favg2, 0.630 Favg3, and 0.784 accuracy, ranking 11th on the official leaderboard. The results show that prompted GPT-4o transfers more effectively to the related-target setting, while additional prompt complexity does not always lead to further improvement.

StanceEval-2026 tests Arabic stance classification under target shift, with evaluation targets absent from the labelled task data. We test two ways to supply target information at inference time. For Track 1, we retrieve eight labelled Women Driving demonstrations from the external ArabicStanceX corpus and score the three label tokens with Gemma-4-31B-IT. For the unseen-target track, we prepend cached target context prompted to be factual and non-evaluative, then fuse label predictions and label-token scores from multiple LLMs, with final assignments constrained by published class totals. The systems obtain 85.22 and 91.78 F_avg2 on Tracks 1 and 2, ranking sixth and fourth in the final standings. On held-out ArabicStanceX tweets, direct prompting scores 79.45 F_avg2; eight random Women Driving demonstrations raise it to 84.78, and similarity retrieval reaches 89.04. Across three target-held-out Mawqif evaluations, a fixed target card raises Gemma-4-31B-IT from 86.54 to 86.93, whereas QLoRA lowers Gemma-4-12B-IT from 84.43 to 83.77. In these separate public evaluations, same-target demonstrations and retrieval improve performance, factual target context provides small, target-dependent gains, and the tested source-only adapter does not improve.

We describe the ZHR system for both tracks of StanceEval-2026, Arabic stance detection on Mawqif-XT, where Track 1 evaluates on seen targets and Track 2 on unseen ones. The system combines fine-tuned Arabic encoders, a few-shot classifier that retrieves labelled examples for a served instruction model, and generative models adapted with LoRA, and turns their blended probabilities into labels with a threshold rule that abstains to None unless one stance class is probable enough, with thresholds fixed on the development split. We report the whole sequence of experiments, including eight that lowered F_avg2 and were not adopted. The main finding is that transfer to a new target is carried by the demonstrations, the labelled examples retrieved into the prompt, rather than by the model itself. Generated on-target tweets in the example pool gave the largest gain on both tracks, matching the surface form of the test tweets helped where the target has a distinctive style, and doubling the examples hurt. Our official results are F_avg2 0.896 with accuracy 0.855 on Track 1, and F_avg2 0.940 with accuracy 0.922 on Track 2. The strongest official baselines reach 0.8425 on the Track 1 development set and 0.7475 on the Track 2 test set.

We present an overview of the **ImageEval 2026** shared task on culturally grounded Arabic multimodal evaluation. It includes two tasks: (i) **AynVQA**, covering spoken visual question answering and image-grounded hallucination detection in English and Modern Standard Arabic (MSA), and (ii) **CRAI-Bench**, evaluating the cultural accuracy of text-to-image generation. A total of 14 teams participated in the test phase, with 12 teams submitting system description papers. Participating systems used a range of approaches, including zero-shot prompting, fine-tuning of vision-language models, speech-recognition pipelines, ensembling, and score calibration. We describe the task setup, datasets, evaluation procedure, and participating systems, and summarize the main results across the different tracks. All datasets and evaluation scripts from the shared task are released to the research community. The shared task highlights challenges in culturally grounded multimodal evaluation, particularly for Arabic speech and image-text reasoning.

We describe the Dynamos submission to ImageEval 2026 Task 1b, image-grounded hallucination detection: given an image and three statements, exactly one of which is grounded, a system must label each true or false. The official ranking metric, Contrastive Instability (CI, lower is better), reduces exactly to one minus the accuracy of picking the grounded statement once a system always outputs a single true label, which removes an entire failure mode (marking zero or several statements true) at no cost. Building on this reduction, we score each statement with an open vision-language model and always output the highest-scoring one as true. Starting from zero-shot Qwen2.5-VL, we low-rank fine-tune the model with a selection objective, train multiple random-seed adapters, and average their scores. Our main finding is that the strongest single model is not the strongest ensemble member: a 32B adapter trained on more data performs substantially better on the labelled devtest split but generalises worse than smaller 7B adapters to the blind test set, yet combining the two model families improves over either pure-family ensemble on both languages, despite the 32B family generalising worse on its own. Our final submitted systems reach CI 0.033 in English (2nd of 8 teams on the official leaderboard) and 0.056 in Modern Standard Arabic (3rd of 6 teams), both well below our zero-shot baselines of 0.058 (English) and 0.110 (MSA).

The increasing adoption of multimodal large language models (MLLMs) has enabled more natural interaction with visual and spoken information, creating new opportunities for intelligent audio-visual question answering (AVQA) and hallucination detection systems. AVQA extends traditional visual question answering by requiring models to jointly understand an input image and a spoken question to generate accurate, context-aware responses. Despite recent advances, AVQA and hallucination detection tasks remain challenging for modern standard Arabic (MSA) due to the limited availability of multimodal datasets and the need to reason across Arabic speech and text modalities alongside visual inputs. In this paper, we present an MLLM-based framework for AVQA that takes an image and a spoken question as input and generates answers. In addition, we utilized this framework to detect hallucination in text-image pairs. The proposed approach is evaluated on the ImageEval 2026 Task (Ayn-VQA) benchmark, which focuses on culturally grounded Arabic AVQA. Experimental results demonstrate that the proposed framework effectively understands spoken queries, grounds their responses in the visual content, and produces accurate answers. The results underscore the capacity and limitations of different LLM architectures for both English and MSA and under various conditions.

Visual hallucination detection in culturally grounded images has remained challenging due to fine-grained visual distinctions and knowledge-dependent statements. In the ImageEval2026 / ArabicNLP26 AynVQA Shared Task, we have participated in Task 1b by developing a system based on Qwen2.5-VL-7B Instruct with 4-bit quantization. We have reformulated hallucination detection as a single choice selection problem to enforce the exactly one-correct-statement constraint and have introduced confidence-based selective refinement using paraphrase ensemble voting. Our final system has achieved 0.942 accuracy, 0.058 CI, 0.000 CFHR,0.942 Q+accuracy and 0.971 Q−accuracy, ranking 6th in the English track. Error analysis has revealed that the remaining failures have mostly stemmed from fine-grained cultural and visual grounding challenges rather than prediction inconsistency.

We present the Team Falcons system for the ImageEval 2026 Task 1b English track. The task provides culturally grounded images from the OASIS/EverydayMMQA resource family and three English statements per image, exactly one of which is visually supported(CITATION). Rather than treating the statements as independent binary decisions, we exploit this structural constraint and formulate each item as image-conditioned three-way classification. In Stage 1, we fine-tune Qwen3-VL-8B(CITATION) with 4-bit QLoRA(CITATION) and estimate uncertainty from the first-token probabilities assigned to A, B, and C. Stage 2 we send only examples whose top-two probability margin is below 0.90 to Gemini 3.6 Flash and GPT-5.4-mini. Their answer replaces the Stage 1 prediction only when both verifiers return the same valid option. On the official test set, Stage 1 obtains a Contrastive Instability (CI) of 0.0360 and Combined Accuracy of 0.9640. Consensus verification improves these scores to 0.0290 and 0.9710, respectively, placing our submission first on the Task 1b English leaderboard. After test labels were released, we found that the verifiers agreed on 91 of 138 routed items and changed 17 predictions: 12 corrections and 5 harmful revisions. These results show that structure-aware fine-tuning and conservative uncertainty routing can effectively combine a local open-weight VLM with proprietary multimodal verifiers.

We describe our systems for both ImageEval 2026 shared tasks. On Subtask 1a (spoken visual QA) we reach 0.962 on English and 0.834 on Modern Standard Arabic against baselines of 0.664 and 0.398, ranking first on the English track and second on MSA, while on Task 2 (CRAI-Bench) our judge ensemble reaches a test Spearman of 0.673 against 0.533 for our reproduction of the organizer baseline. Three effects account for most of these margins: answer-token bias (digit labels suppress the answer 0, and ordinal words are worth 0.35 accuracy), Arabic speech recognition (isolated from visual reasoning by the parallel track design), and tie-breaking (calibration collapses ranks and a second judge separates them, while error correlation of 0.89 rules out cancellation). A post-submission study on degraded audio finds four transcript-level interventions null and measures a 9% answer-flip rate between identical runs at temperature 0, a noise floor for single-run ablations.

We describe the VisionEval submissions to ImageEval 2026, covering Task 1b in the English and MSA tracks and Task 2 (CRAI-Bench). For Task 1b the three statements of an item share a stem and exactly one is grounded, so we replace the three independent True/False calls of the released baseline with a single three-way forced choice, read from the logits of the candidate digit tokens. The output then satisfies the one-true constraint by construction, and contrastive instability falls from the baseline 0.2674 to 0.0670 in English and from 0.4277 to 0.1010 in MSA. The organisers’ baseline is Qwen2.5-VL-3B prompted one statement at a time, and it identifies the grounded statement about as often as our 7B system does in English (0.9310 against 0.9330), so most of that gap is the output format rather than perception or scale. Task 2 is handled by a six-parameter model over the caption version, which placed first of five teams at Spearman 0.8258; an analysis of what that model reads shows the version field is recoverable from the released instance identifiers, so the result characterises the benchmark more than the system. We also show that the five reported Task 1b columns collapse to one degree of freedom for any system marking one statement true per item, an identity satisfied exactly by all fourteen rows of the final standings.

Large multimodal models tend to hallucinate visual detail fluently, which limits their deployment for fine-grained interpretation. We present HALDETECT, our system for the English hallucination-detection track (Task 1b) of ImageEval 2026, in which a system must identify, from an image and three culturally plausible statements, the single visually grounded one. We frame the item as one contrastive decision, emit the answer before its explanation, and structure reasoning around colour/texture, shape/form, and context. Our best submitted adapter fine-tunes Qwen2.5-VL-7B-Instruct with 4-bit QLoRA while freezing the vision encoder and reaches Contrastive Instability (CI) 0.035 on the 1,000-item test set; we placed third of eight teams. Development experiments show that answer order can matter more than model scale and that adaptation beats prompting alone. Retrospective paired analysis of the released gold labels confirms the QLoRA gain over the best prompt but not the small gap between the devtest-selected and best-test adapters, and reseeding all four training sizes shows that the apparent data-scaling curve does not survive a seed change. The 35 residual errors are culturally plausible function, material, and recognition distinctions; naive adapter voting does not help.

We describe our submission to Subtask 1b of ImageEval 2026. The task consists of selecting the single visually valid statement from three candidates paired with an image, and is in both English and Modern Standard Arabic (MSA). Our system uses a pre-trained 7B vision–language model in a zero-shot setting, without task-specific training or fine-tuning. We formulate the task as a constrained selection problem and choose the answer using next-token log-probabilities. To reduce answer-position bias, we apply a Latin-square rotation to the candidate statements. Our system achieves CI scores of 0.0510 on the English test set and 0.0960 on the MSA test set, reducing CI relative to the official 3B baselines reported on devtest by 83.7% and 80.4%, respectively.

We present CUET_InferX, our submission to Subtask 1a (Spoken VQA, English-audio) of the ImageEval 2026 shared task on the AynVQA dataset. The task requires predicting the correct answer option solely from an image and a raw audio clip containing both a spoken question and three candidate options. We fine-tune Qwen2.5-Omni-3B with LoRA applied exclusively to the language reasoning backbone, keeping both audio and vision encoders frozen, under a single 16 GB Kaggle T4 GPU. Training uses a time-budgeted, class-stratified subsample of 2,055 examples. Our system achieves 91.20% test accuracy (Balanced Accuracy 87.72%, Macro-F1 80.50%), placing 3rd position, with gains of +7.20 accuracy and +15.95 Macro-F1 over the zero-shot base. We compare against a cascaded ASR→VLM alternative and zero-shot 3B/7B baselines, and explain with devtest evidence why we submitted the end-to-end fine-tuned model despite lower raw test accuracy than the cascade.

Across two ImageEval 2026 tracks, our best results came from analyzing the evaluation machinery rather than from a stronger model, yielding 1st place on hallucination detection (Task 1b, Modern Standard Arabic) and 4th on CRAI-Bench (Task 2). On CRAI-Bench, which asks systems to predict human cultural-accuracy scores for AI-generated images of Qatari scenes, a blind frequency prior that never looks at an image reaches dev Spearman 0.6878, above the organizers’ GPT-4-as-judge baseline of 0.625; adding two vision judges, with blend weights tuned against the global ranking metric rather than by scene-level cross-validation, reaches 0.7806 on the blind test. The weighting correction alone is worth +0.0296 test Spearman, because scene-level cross-validation is structurally blind to the within-cell ranking that judges contribute. On Task 1b, reframing "label three statements true/false" as "select the one grounded statement" makes invalid labelings structurally impossible and reaches 0.964 combined accuracy on the blind test, against an official baseline of 0.508. A matched-model ablation shows most of that headline gap is the underlying judge, but the constraint still adds +0.026 combined accuracy and eliminates the 4.0% of items that otherwise receive structurally invalid labelings; meanwhile upgrading to a stronger judge is worth nothing measurable and a three-model vote scored lower.

We present a reference-based evaluation framework for CRAI-Bench, Task 2 of ImageEval 2026. The method uses GPT-5.4 to extract cultural evidence from the reference image, caption, and generated image. This evidence is converted into metric-specific numerical features and mapped to CRAI scores using a small regressor trained on the released data. On the development set, the proposed approach improves composite Spearman correlation over direct GPT-5.4 scoring, with strong gains on certain metrics. We also analyze performance across caption versions and show that evaluation becomes less reliable for some dimensions as captions become more generic.

We describe the Digilians submission to Task 1a (Spoken Visual Question Answering, MSA track) of the ImageEval 2026 shared task (Abdaljalil et al., 2026). Our system splits the problem into independently cacheable stages: speech recognition, schema-constrained transcript parsing that recovers a question and three candidate answers from continuous Arabic speech, a conditional repair mechanism that escalates to a larger ASR model only for semantically degenerate parses, and a joint multiplechoice visual reasoning stage that evaluates all three candidates against the image simultaneously rather than scoring them independently. On the labeled development split (500 records, synthetic voice-cloned audio), our production configuration reaches 79.0% accuracy against a 39.8% official baseline, with nearly the entire gain over an earlier configuration coming from a single change to the answer-selection model. On the blind final test split (1,000 records, human-recorded audio), the same frozen configuration scores 65.6% accuracy / 57.4% balanced accuracy / 50.4% macro F1, a substantial drop we attribute to the dev-to-test recordingcondition shift and report as our most important finding. We also surface a caching-related reproducibility issue uncovered during development and discuss what both findings imply for pipelines built the same way.

Dialectal Arabic machine translation (MT) remains challenging despite recent progress in Arabic language technologies, particularly because effective translation requires modeling not only semantic content but also dialectal variation, conversational context, speaker and addressee characteristics, and sociolinguistic appropriateness. Moreover, conventional MT metrics provide limited insight into the linguistic errors produced by dialectal systems. We present the AlexandriaX 2026 Shared Task on Dialectal Arabic MT, which addresses these challenges through three complementary subtasks: (1) context-aware English-to-Dialectal Arabic dialogue translation across 13 Arabic varieties, (2) cross-dialect Arabic translation in the financial domain covering six Arabic dialects, and (3) span-level MT error detection and classification using linguistically motivated error categories across five Arabic varieties. The shared task attracted 38 registrations for Subtask 1, 33 for Subtask 2, and 35 for Subtask 3. Twelve unique teams submitted their system description papers, all of which we accepted for publication. The best system on Subtask 1 achieved 30.42 spBLEU in the constrained track and 33.49 spBLEU in the unconstrained track. On Subtask 2, the top system obtained 28.40 spBLEU. On Subtask 3, the best system achieved an overall score of 49.82, outperforming the 24.29 baseline. Taken together, the results of the leading systems across the three subtasks highlight the benefits of explicit dialect modeling, context-aware generation, retrieval and reranking, and specialized approaches to interpretable MT error analysis. All the resources of AlexandriaX 2026 shared task are publicly available, including data, baselines, and evaluation code on our project page: https://alexandriax.dlnlp.ai.

We present TTLab’s submission to the AlexandriaX-2026 Subtask 3 on Arabic MT error span detection and classification. Our system frames the task as token-level classification over surface forms, preserving character offsets to ensure exact alignment with the evaluation metric. To handle severe label imbalance, we employ a focal loss with class weighting and dialect-specific decoding thresholds. Among six Arabic pre-trained encoders, MARBERTv2 achieves the best overall performance of 40.8 and 40.91 on the development and test set, respectively, ranking 3rd out of all participating teams. While our system localizes error spans effectively, classification of rare error types remains challenging, highlighting the need for data augmentation for tail categories.

This paper describes the Rosetta system for Subtask 1 (Context-Aware English-to-Dialectal Arabic Dialogue Translation) of the AlexandriaX shared task, participating in both constrained and unconstrained tracks. The approach fine-tunes a LoRA adapter on NileChat-3B using structured system/user prompts that condition generation on dialect and dialogue context. For the unconstrained track, the adapter is additionally pretrained on MADAR and PADIC. Rosetta ranked 4th in the constrained track (spBLEU 26.10) and 5th in the unconstrained track (spBLEU 25.09). The experimental results demonstrate that external pretraining helps only two of thirteen dialects while slightly hurting overall performance, suggesting negative transfer.

This paper describes the "Axiom" team’s submission to the AlexandriaX-2026 Shared Task, focusing on Subtask 3: MT Error Detection and Classification. We frame the task of predicting fine-grained error spans and their corresponding LQM categories as a joint token classification problem using a BIO tagging scheme. Evaluating multiple Arabic encoders, we find that dialect-specific pretraining significantly outperforms generative extraction baselines. To address severe label imbalance across error categories, we apply a square-root scaled inverse-frequency class-weighted loss. Furthermore, we introduce a confidence-thresholded decoding strategy at inference to reduce false-positive span over-prediction. Our final system achieved highly competitive performance, ranking 5th on the official leaderboard. We highlight that while overlap-based detection is robust, exact span matching remains a persistent challenge due to the non-standardized nature of dialectal Arabic orthography.

We describe the Charles University (CUNI) submission to the AlexandriaX-2026 Subtask 1, a shared task on English-to-Arabic dialectal conversation translation covering 13 Arabic varieties. Our approach fine-tunes Gemma-4-31B-it with QLoRA on the Alexandria training data at the conversation level, using rich contextual prompts that include domain metadata, participant names and speaker labels. We explore multiple training data mixes and hyperparameters, yielding a pool of candidate translations for the test set. These are reranked with a genetic-algorithm-enhanced minimum Bayes risk (GA-MBR) decoding using sentence-level spBLEU as the fitness function. Fine-tuning improves over the unmodified base model by approximately 5.2 ChrF points on the development set (from 45.2 to 50.4 ChrF for the best single checkpoint, or 25.1 to 32.6 spBLEU); GA-MBR reranking over a 40-candidate development pool yields 51.3 ChrF / 34.0 spBLEU, a further improvement of 0.9 ChrF points over the best pool member. Our submission ranked second in the unconstrained track.

Arabic encompasses diverse spoken varieties with limited parallel data, making both translation and automatic evaluation difficult beyond Modern Standard Arabic. We present the NAMAA Community submissions to all three AlexandriaX-2026 subtasks: English-to-dialect dialogue translation, cross-dialect translation, and dialectal MT error detection. Our systems combine few-shot prompting, fine-tuned encoder-decoder and decoder-only models, and encoder- and LLM-based span detection. Held-out ablations show that minimum-Bayes-risk consensus can suffer from model-family imbalance; we therefore use per-dialect routing for translation and 2-of-3 character-level agreement for error-span detection. Our systems rank fourth and fifth of six in the unconstrained and constrained Subtask 1 tracks, third of four on Subtask 2, and second of six on Subtask 3. The Subtask 3 system achieves 0.5448 overlap-F1, exceeding the overall winner on span localization, with category assignment remaining the main weakness. Code and resources are publicly available.[<https://huggingface.co/collections/NAMAA-Space/namaa-at-alexandriax-2026>]

Dialectal Arabic MT error detection asks a sys tem to locate error spans in a machine trans lation and label each with a linguistic error category. It is hard because the dialects lack a standardised orthography and the error cate gories are severely imbalanced. In Subtask 3 of the AlexandriaX 2026 shared task, the in put is an English sentence and its machine translation into one of five Arabic dialects, and the output is character-level error spans with their LQM categories. The central difficulty is collapse onto the majority sociolinguistics class, 58% of training spans, leaving rare cat egories undetected. To address this, we fine tune NileChat-3B-Base with QLoRA on a sin gle 8 GB consumer GPU, casting the task as schema-prompted JSON generation and com bining character-position annotations that let the model copy offsets rather than count them, a category guide with one worked example per error type, and self-consistency decoding merging five samples by character-level union voting. Our system scores 0.3484 Overall on the official test set. We report two negative results: early stopping on validation loss de grades performance, since rare categories are acquired only in later epochs, and fine-tuning on the skewed distribution lowers macro-F1 be low the un-fine-tuned baseline until the cate gory guide is added

Cross-dialect Arabic machine translation asks for translation between Modern Standard Arabic and regional dialects with almost no parallel text. We present a system for Subtask 2 of AlexandriaX-2026, built on two findings about where the gains come from. First, data was not the binding constraint: once our mined-and-synthetic recipe had decelerated to +1.8% relative spBLEU per round, a single increase in LoRA capacity beat three further data-tuning rounds combined (+8.9%). Second, planned epoch budgets were consistently too short, and training past them was the largest single lever at every stage. We fine-tune NLLB-200-distilled-600M on a three-tier corpus of real, mined and synthetic pairs, then refine it with Direct Preference Optimization. Ablating three reward designs isolates a third finding: rewarding agreement with self-distilled references regresses real quality despite healthy training dynamics. Restricting the reward to references we trust, with a reference-free substitute for the two zero-resource dialects, yields a real gain that a further iteration gives back. Every configuration was scored on the official leaderboard. Our submission reached spBLEU 25.84 and chrF++ 39.02, ranking 2nd of 4 teams; post-evaluation experimentation reached 28.04.

We describe REGLAT’s submission to AlexandriaX-2026 Subtask 3, which requires character-offset error spans in dialectal Arabic machine translations and a Linguistically Motivated Multidimensional Quality Metrics (LQM) category for each span. The system separates localization from classification. Four CAMeLBERT-DA and MARBERTv2 BIO taggers project token probabilities to a shared character grid; whitespace-segment probabilities are then thresholded, bridged, and optionally dilated. A second four-member ensemble classifies detected spans using the marked hypothesis, English source, and translation direction, with class weighting and fitted logit biases. REGLAT ranked fourth out of six teams with an official Overall Score of 0.4021 (Overlap F1 0.4272; Error-Class score 0.3770). Its Exact Match F1 of 0.0041 exposes severe boundary imprecision. We document a scorer mismatch in development-time selection and therefore treat earlier development values as proxy-metric results, not official-metric estimates.

Translating English dialogue into dialectal Arabic depends on the target dialect, conversational context and speaker attributes, and these challenges are compounded by the scarcity of dialectal parallel data. We present our system for the AlexandriaX-2026 Context-Aware English-to-Dialectal Arabic Dialogue Translation constrained track, which restricts participants to the provided Alexandria dataset and models of ≤5B parameters. We fine-tune open LLMs, exploring different context composition strategies for training and inference, and find that mismatched strategies can outperform matched ones depending on the model. We further apply quality-aware decoding (QAD) via Minimum Bayes Risk, tuning the utility function and candidate pool size for consistent gains over greedy decoding. Our final system, a fine-tuned Gemma-3-4B-it with spBLEU-based QAD and a pool size of 50 samples, achieves 26.48 spBLEU and 42.3 chrF++, ranking third on the constrained-track leaderboard.

Dialectal Arabic dialogue translation requires modeling both regional variation and conversational context under uneven supervision. We introduce LahjaMT, a compact system for translating English dialogue into thirteen country-level Arabic varieties. LahjaMT adapts the 3.12B-parameter NileChat-3B-Base model with LoRA, conditions translation on dialogue history and metadata, and routes each dialect to a checkpoint–prompt expert selected on the held-out selection set. The constrained system shares one backbone across all varieties; the unconstrained system changes only the Libyan and Sudanese routes, replacing them with specialists trained on auxiliary data. LahjaMT ranked second in the constrained track with 28.30 spBLEU and third in the unconstrained track with 28.54 spBLEU.

This paper describes team Alkhder’s submission to Subtask 1 of the AlexandriaX-2026 shared task, which asks systems to translate English dialogue turns into thirteen Arabic dialects. We fine-tune NileChat-3B-Base with QLoRA on the Alexandria parallel dialogue corpus, training one set of adapter weights for all dialects. Each turn is translated conditioned on the target country and dialect, the domain, the participant roles, the speaker–addressee gender direction, and the preceding dialogue context. At inference, conversations are translated autoregressively and the model is fed its own previous outputs back as context without using any gold target-side test information. Under the constrained-track, using only Alexandria data, it trains in about one hour and translates the private test set in under two hours. The system reaches a country-macro-averaged spBLEU of 20.36 and chrF++ of 36.64, on both tracks.

We report first-place systems for all four evaluated AlexandriaX-2026 settings: both English-to-dialect translation tracks, cross-dialect translation, and translation-error detection. In unconstrained translation, our minimum Bayes risk (MBR) selection restricts output authority: four local systems can supply the translation, while seven additional systems contribute only reference evidence. Candidate-constrained MBR beats the strongest complete candidate in all 13 dialects and gains 1.2554 macro spBLEU overall, reaching 33.4891. It also surpasses a retrospective oracle selecting the best whole candidate system per dialect. In the constrained track, country-routed LoRA adapters over a shared NileChat-3B-Base backbone incorporate the released Libyan and Sudanese supervision absent from the training and development splits, raising macro spBLEU from 29.5969 to 30.4233. Post-evaluation, a single adapter trained on all released Alexandria data reaches 30.4733 macro spBLEU; averaging model weights across three training seeds raises this to 30.9726 while retaining one model for inference. For cross-dialect translation, nonparallel target-dialect retrieval guides deliberation, raising Gemma-4-31B-IT without fine-tuning from 24.6898 to 28.4011 spBLEU. For translation-error detection, GPT-5.5 proposes semantic error regions, while MARBERTv2 controls their character boundaries and corroborates additional spans, raising Overall from 0.4691 to 0.4982.

We present the findings of the Arabic Sentence Segmentation (AraSEG) Shared Task, organized at the Fourth Arabic Natural Language Processing Conference (ArabicNLP 2026). Given an Arabic document, the task requires systems to identify sentence boundaries and segment the text into coherent sentences. Systems are evaluated on the genre-diverse AraSEG corpus under four input settings that vary in punctuation availability and paragraph structure, each with a Closed Track restricted to the provided training data and an Open Track permitting publicly available external resources. 22 teams from 13 countries registered, 16 submitted systems, and 14 submitted system-description papers. Across the eight subtask-track combinations, the best systems achieve F1 scores ranging from 89.1 to 95.3. Performance declines as segmentation cues are removed, with the largest drop occurring when both punctuation and paragraph structure are unavailable. Pretrained-encoder ensembles and careful prediction calibration emerge as the most effective approaches, while gains from LLMs, external data, and complex architectures vary across systems and settings.

Arabic sentence segmentation without punctuation or paragraph breaks requires predicting boundaries from context. The AraSeg Open Track supplies a parallel punctuated view, but its marks do not consistently indicate sentence endings. We train two CAMeLBERT-MSA ensembles, one for the target No-Punctuation No-Paragraph view and one for the auxiliary No-Paragraph view. A discrete rule combines their boundary decisions with punctuation classes projected onto the target words. Model fitting and selection use only the 174-document official training split: five-fold cross-validation selects the training duration, and out-of-fold predictions calibrate the decoders and fusion rule. Development is reserved for evaluation, and the public test is unused. The system achieves 84.61 document-macro F1 on development and 84.88 on the official blind test, with blind precision of 81.87 and recall of 89.50. It predicts too many boundaries in 80.6% of development documents, although some of its lowest scores result from missed boundaries.

We present NAMAA Community’s systems for the four closed-track Arabic sentence-segmentation subtasks of AraSeg 2026. Heterogeneous ensembles combining LoRA-adapted bidirectional language models, XLM-R variants, a character-level SaT model, and punctuation-aware models outperform their development-selected single-model baselines. Large weights assigned to weaker but architecturally distinct members are consistent with complementary errors, although they do not establish causality. A training-only logistic MEMM improves F1 by +0.9 and +0.6 points on the punctuation-free subtasks but degrades both punctuated ones; the same directional pattern appears on the blind split. Aligning punctuated and punctuation-free inputs reveals F1 points of oracle headroom against the contemporaneous NoPnx locks, while only 52.5% of removed tokens are covered by the restoration label inventory. Genre-proxy analysis localizes the largest degradation to scriptural text. On the blind split, in the closed track, our systems achieved F1 scores of 94.4/91.3/89.9/87.0 on PA, NP, NoPnx-PA, and NoPnx-NP, respectively, ranking 6th/3rd/4th/3rd. The same systems, entered unchanged in the open track, rank 4th/2nd/3rd/3rd. Implementation and experimental resources are publicly available.[<https://github.com/NAMAA-ORG/NAMAA-Community-AraSeg-2026>]

Identifying sentence boundaries in Arabic text is a challenging token classification task due to extreme dialectal variations, the absence of rigid punctuation rules, and document lengths that frequently exceed the capacity of standard transformer models. In this paper, we describe Team Axiom’s submission to the Paragraph-Aware Open track of the AraSeg Shared Task 2026. We propose a highly regularized, dual-model ensemble combining MARBERTv2 and CAMeLBERT-MSA. To address the challenge of excessively long documents, we employ a sliding-window tokenization strategy. Furthermore, we adapt our base models to the task corpus using Domain-Adaptive Pretraining (DAPT). Our final system achieved a Macro-F1 score of 90.2% (Precision: 91.9%, Recall: 89.5%), securing 7th place in the shared task.

We describe the iMak AI Lab submissions to the closed track of the AraSeg Shared Task 2026 on Arabic sentence segmentation, covering all four subtasks: Paragraph-Aware (PA), No-Paragraph (NP), No-Punctuation Paragraph-Aware (NOPNX-PA), and NoPunctuation No-Paragraph (NOPNX-NP). Our system treats segmentation as binary token classification with pretrained Arabic encoders finetuned only on the 128K-token AraSeg training split. The single-model recipe encodes documents in overlapping 512-token windows and trains for three times more epochs than the standard schedule, which is the largest source of improvement in our pipeline. Per-subtask ensembles are then built by greedy forward selection over a pool of seven encoder families, several seeds, and multi-variant joint-training variants. Decoding applies deterministic structural constraints and a decision threshold tuned on the development set; no open-test labels or leaderboard feedback enter any modeling decision. The systems reach blind-test F1 scores of 94.1 (PA), 91.0 (NP), 89.5 (NOPNX-PA), and 85.8 (NOPNX-NP), above the official baseline on all four subtasks. We report ablations for each design decision, including several that did not help.

We describe NLPBeasts’ submission to the Paragraph-Aware (PA) Closed track of the AraSeg 2026 shared task, which frames Arabic sentence segmentation as per-token binary boundary classification. We compare two closed-track-compliant systems trained with out any pretrained embeddings or language models: (1) a feature-based CRF using handcrafted punctuation, brackets, and paragraphs distance features, and (2) a character- and word-level BiLSTM with a linear-chain CRF decoder, both trained entirely from scratch on the shared task’s training and development splits. The feature-based CRF plateaus at a hard recall ceiling (≈82–83%) regardless of regularization strength, while the BiLSTM-CRF captures substantially more boundary cases through learned context. Our BiLSTM-CRF system, using only permitted training data with the development split reserved strictly for model selection, achieves P=86.37, R=90.51, and F1=87.60 on the blind test set. We additionally report ablations on chunk length and training hardware (CPU vs. GPU), neither of which meaningfully changes development F1, suggesting our system’s performance ceiling on this closed-track data size is data-limited rather than compute- or context-length-limited.

We describe our submission to the AraSeg 2026 shared task on Arabic sentence segmentation (Paragraph-Aware, Closed track). We cast segmentation as binary token-boundary classification and build a system around three ideas: (i) task-adaptive pretraining (TAPT) of encoders on the task’s own training text, (ii) regularized fine-tuning with R-Drop and boundary smoothing, and (iii) a hand-designed cross-family ensemble of architecturally diverse Arabic encoders. Our final system reaches a boundary F1 of 94.8 on the blind test set (ranking 3rd of 12 teams, 0.5 points from the top-ranked team). Beyond the score, we report a practical finding backed by a series of blind submissions: on a small development set, a compact diversity-first ensemble generalizes better to unseen data than ensembles selected by exhaustively maximizing development F1, and TAPT helps only under-adapted backbones while hurting a heavily-pretrained one. We also report five techniques that did not help, to save future practitioners the effort. Our code is publicly available at https://github.com/Body123/araseg-2026-knights-st.

We describe ENSAR, our submission to the AraSeg 2026 shared task on Arabic sentence segmentation, where a system must decide, for every word in a document, whether a sentence ends there. The four tracks strip punctuation, paragraph breaks, both, or neither. We start from an ensemble of five fine-tuned encoders decoded with a length-aware semi-Markov program. It is strong, but most of its mistakes are places where all five models are wrong and all five assign high confidence, so averaging them or moving a threshold changes nothing. ENSAR adds a reasoning stage on top. Two LLM juries, each holding a written policy learned from its own graded errors on the training split, debate every document and apply only the edits both endorse. Our system ranked first on three of the four tracks in both the closed and open settings.

Sentence boundary detection in Arabic is inherently challenging due to ambiguous orthographic punctuation conventions, prolific coordinating conjunction clitics, and extensive stylistic variability across literary, religious, news, and web domains. In this paper, we present our system submission to the Closed Track of the AraSeg 2026 Shared Task. We formulate sentence boundary detection as a token-level sequence classification problem utilizing an ensemble of fine-tuned Arabic Transformer encoders: AraBERTv02, CAMeLBERT-MSA, and ARBERTv2. Our pipeline integrates an alignment-safe normalization strategy that strictly preserves gold token spans, overlapping sliding-window logit margin aggregation to counteract chunk-boundary degradation, stochastic punctuation masking during training, and calibrated positive-class loss weighting. In the official blind evaluation across eight diverse genres, our framework demonstrated strong boundary recovery across all four task tracks, securing 2nd place in both the Paragraph-Aware setting (F₁ = 95.0%) and the No-Paragraph setting (F₁ = 91.3%).

The AUC DL Lab system addresses the AraSeg 2026 shared task on Arabic sentence segmentation across four conditions: paragraph-aware (PA), no-paragraph (NP), no-punctuation paragraph-aware (NOPNX-PA), and no-punctuation no-paragraph (NOPNXNP), in both the closed and open tracks. We cast segmentation as imbalanced word-level boundary classification and pair fine-tuned Arabic encoders with two complementary architectures: HDP-GRAPH, which models syntactic structure via dependency-graph message passing, and a quantized decoder-only LLM using Just Pass Twice (JPT) with QLoRA for contextual boundary classification. Predictions are calibrated and combined through differential-evolution ensemble weighting. On the blind test set, our closed track systems reach 94.5, 90.8, 89.2, and 86.0 F1 (ranking 5th, 6th, 6th, 7th), and our open-track systems reach 94.6, 90.9, 89.7, and 86.0 F1 (ranking 3rd, 4th, 4th, 5th) on PA, NP, NOPNX-PA, and NOPNX-NP, respectively. Architectural diversity grows more valuable as punctuation and paragraph cues are removed, while external data yields only marginal gains once in-domain coverage is adequate.

Arabic sentence segmentation becomes substantially more difficult when punctuation and paragraph boundaries are unavailable. The AraSeg Shared Task 2026 provides a controlled benchmark for studying this problem under four conditions that progressively remove structural information. We present the Wathiq-SPU submission to all four closed-track conditions: Paragraph-Aware (PA), No-Punctuation Paragraph-Aware (NoPnx-PA), No-Paragraph (NP), and No-Punctuation No-Paragraph (NoPnx-NP). We formulate sentence segmentation as token-level binary boundary classification using pretrained Arabic Transformer encoders and combine calibrated predictions with track-specific ensemble strategies that adapt to the amount of available structural information. CAMeLBERT-MSA serves as our primary encoder, while the most challenging NoPnx-NP condition additionally benefits from AraELECTRA-based backbone diversity. Our experiments reveal that the most effective source of model diversity depends on task difficulty: diversity among independently trained models provides robust gains when some structural information is available, whereas backbone diversity becomes increasingly important when both punctuation and paragraph cues are removed. The selected systems achieve F1 scores of 93.4, 87.9, 90.0, and 84.8 on the PA, NoPnx-PA, NP, and NoPnx-NP tracks, respectively

Arabic sentence segmentation without punctuation or paragraph boundaries requires models to infer boundaries only from lexical, syntactic, and discourse context. We describe FAAMG’s submission in the closed track of AraSeg 2026 Subtask 4 (NoPnx-NP), restricted to only 174 training documents for fine-tuning. Our system is a weighted ensemble of eight models based on AraBERT, SaT, and CAMeLBERT BiLSTM-CRF. The ensemble covers a variety of regularization, encoder capacity, and sequence modeling. The resulting system achieves 86.7 document-macro F1 on the Blind Test set ranking fourth on the leaderboard. Through leave-one-out ablations we show that the ensemble contributes complementary information over the individual models.

We describe our submissions to the four subtasks of the AraSeg 2026 shared task in both tracks. Our components are conventional word-level boundary taggers, and our contribution is a measurement of what combining them is worth. Averaging 21 members gains +0.89 F1 over the best single model, and replacing the uniform probability average with a learned logistic combiner adds a further +0.23 (+1.12 in total), reaching significance only on the two subtasks that lack punctuation. Both are smaller than the encoder swap (+1.82). What the combiner buys is freedom from curation. Added to the base encoder, each of six different pretrained encoders helps under the combiner (+0.12 to +0.72, mean +0.34), whereas under averaging three of the six are harmful, by as much as 0.62. Another seed of an existing encoder is worth nothing under either rule; members should therefore be diversified rather than curated, provided the weights are learned. We also report three null results for inference-time structure, including a semi-Markov CRF whose failure we trace to a length prior that is uninformative at corpus level. Our closed-track systems placed fifth and seventh where ranked. In the open track, pretraining on 3.2M tokens of silver boundaries from Arabic Wikipedia before fine-tuning on the gold training split is worth +0.58 F1, and two of the closed-track findings reproduce on that external data.

This paper describes the Abjad AI team’s sub- mission to the AraSeg 2026 Shared Task on Arabic sentence segmentation, covering the two punctuated tracks: Paragraph-Aware (PA) and No-Paragraph (NP). Rather than substituting a larger encoder, we retain the CAMeLBERT- MSA reference system and ask how much head- room remains in the task’s formulation. Our system combines four components: boundary labels aligned to each word’s final sub-token, an auxiliary classification head on the penul- timate encoder layer trained jointly with the final-layer head, class-weighted cross-entropy for boundary sparsity, and overlap-averaged decoding that scores documents of arbitrary length without truncation. We ablate each com- ponent in isolation, including the two that the reference system already uses. Our solution achieves F1 scores of 91.7 and 89.7 in the blind tests of the PA and NP sub-tasks, respectively.

We submit systems to both the closed and open divisions of all four tracks of the AraSeg 2026 shared task (PA, NP, NoPnx-PA, NoPnx-NP): a supervised 3-backbone ensemble (CAMeLBERT-MSA, AraBERTv02, and AraELECTRA, each with a gated-residual head, R-Drop regularization, and stochastic weight averaging), and a zero-shot LLM system prompted with explicit, track-specific segmentation rules instead of labeled examples. The ensemble uses no external resources, so it is entered in both divisions everywhere; for NoPnx-NP specifically, where every structural cue is stripped away, we additionally blend its predictions with the LLM system’s, in a real-data-only variant (closed-eligible) and a wiki-augmented variant trained with added pseudo-labeled Wikipedia data (open-only). On the officially-scored blind test, our best submission per track reaches 94.70 F₁ (PA), 91.00 (NP), 91.20 (NoPnx-PA), and 87.40 (NoPnx-NP). The supervised ensemble is the stronger, more consistent choice overall; the one exception is NoPnx-PA, where our LLM system’s unmodified seed prompt outperforms it and combining the two systems does not help—the only track where that happens. We report in full the structural annotation ambiguities we uncovered while building the LLM system.

We describe the main findings of the second IslamicEval shared task on hallucination detection in LLM-generated Islam related text (IslamicEval 2026). While the first occurrence of the task in 2025 separately evaluated hallucination detection/correction in quoted Quranic verses and Hadiths alongside a distinct QA retrieval subtask, this year integrates these objectives into four fine-grained subtasks with a much larger provided training set. Given an Islamic question and an LLM response, systems must: 1) detect all cited Quran and Hadith references (including verse numbers and narrators); 2) identify hallucinated citations; 3) correct the hallucinated content; and 4) assess the relevance of every citation to the prompt. Sixteen teams participated in the final phase across the four subtasks, utilizing diverse pipelines that highlight both the challenge of the task and the necessity of mitigation strategies in Islamic AI applications.

This paper describes our submission to the IslamicEval 2026 Task 1 (Span Detection) shared task, which aims to identify and classify evidence spans in AI-generated Arabic Islamic answers. We formulate the task as token-level sequence labeling and fine-tune CAMeLBERT-Mix using BIO tagging, overlapping-window inference, and prediction reconciliation to detect four evidence categories: Ayah, matn, isnad, and claimed_source. On the official development set, our approach achieves a Precision of 95.55%, Recall of 97.06%, F1-score of 96.30%, and Token Accuracy of 98.66%. Multi-seed and qualitative analyses demonstrate the robustness of the proposed framework for Arabic evidence span detection.

Large Language Models (LLMs) have significantly improved general-purpose question answering but remain prone to hallucinations in fact-sensitive domains such as religion. IslamicEval 2026 addresses these challenges by providing a fine-grained benchmark for evaluating hallucination and relevance in Arabic Islamic citations. This paper presents our approaches to three tasks: Citation Span Detection (Subtask 1), Hallucination Identification (Subtask 2), and Answer Relevance (Subtask 4). For Citation Span Detection, we formulate the problem as BIO-based token classification and fine-tune AraBERT using overlapping sliding-window encoding to process long responses effectively. For Hallucination Identification, we introduce a segment-level verification framework that augments each citation with adaptively extracted contextual evidence and citation-type information. For Answer Relevance, we combine hybrid lexical-semantic retrieval over authenticated Qur’anic and Hadith corpora with Transformer-based classification, enriching retrieved evidence with citation-specific context such as neighboring Ayat and explicit Surah–verse references. On the official test set, the proposed solutions achieved an F1 of 0.9725, an accuracy of 86.62%, and a macro-F1 of 0.7721 for Subtasks 1, 2, and 4, respectively, demonstrating the effectiveness of the proposed approaches.

IslamicEval 2026 Subtask 4 asks whether a Qur’anic or Hadith citation is useful for answering an Arabic question, rather than merely related to its topic. We treat this decision as ranking under shared response context. An Arabic instruction prompt is scored by the first answer-token logit difference between na’am ("yes") and la ("no"). We adapt ALLaM 7B with LoRA and a response-balanced RankNet loss over 1,063 positive-negative pairs drawn from 348 mixed-label responses. At inference, the model uses deterministic logits and one calibrated threshold; it does not generate or sample an answer. On Development, the frozen checkpoint obtained 0.867761 Question Macro-F1 under question-grouped out-of-fold threshold calibration, compared with 0.777935 for zero-shot ALLaM. Our primary submission, AhmedNLP_AllamR, scored 0.767880 on the official test and was the best of our three submitted runs. The largest Development-to-test drop occurred on all-negative questions. This result suggests retaining pairwise supervision while adding direct supervision for homogeneous negative groups.

Large language models are increasingly used to answer questions about religious texts, yet their fluent responses may contain misquoted, misattributed, or fabricated citations. Detecting such hallucinations requires identifying citation components, verifying their correctness, and recovering the canonical text when a quotation is incorrect. We present Yaqeen’s submission to the IslamicEval 2026 shared task, which evaluates these capabilities through three subtasks: detection of Qur’anic verses, Hadith text, chains of narration, and claimed sources; identification of incorrect citation fragments; and correction of erroneous Qur’anic and Hadith quotations. We develop task-specific systems combining AraBERT-based character-level sequence labeling with retrieval-based matching, classification, and reranking. Our systems achieve 96.78% character-level Macro-F1 on Task 1, 91.07% accuracy on Task 2, and 78.02% accuracy on Task 3. The results show that combining contextual Arabic representations with retrieval-based evidence is effective for fine-grained hallucination detection and correction in Arabic Islamic content.

This paper presents our participation in IslamicEval 2026 Subtask 4, citation relevance classification for Qur’an and Hadith spans. We propose two independent systems for this task. The first is a generative large language model prompted to produce a structured, step by step relevance judgment. The second is a domain pretrained sentence encoder fine tuned with triplet loss on citation relevance data. Our evaluation compares two generative backends and three encoders pretrained on different Arabic varieties, using macro averaged F1 as the metric. Results show on development set that chain of thought judgment with ALLaM-7B consistently outperforms our embedding similarity ensemble, reaching 0.774 macro F1 against 0.357 for the ensemble, and outperforms an equivalent Fanar-S-1-7B system on seven of eight prompting strategies. Additionally, AraBERT-v02, pretrained on Modern Standard Arabic, is the consistently strongest encoder, outperforming Classical Arabic pretrained encoders in this citation relevance task. Our best submitted system, ALLaM with few shot prompting without chain of thought, scores 0.748 macro F1 on the blind test set, narrowly ahead of the same backend with chain of thought added, at 0.744.

We present SanadAI, a modular system for the IslamicEval 2026 shared task on fine-grained hallucination detection and correction in Arabic Islamic content. We participate in Subtasks 1 and 2, which address citation span detection and hallucination detection, respectively. For Subtask 1, we fine-tune AraModernBERT as a BIO sequence tagger to identify Quranic verses ayah, Hadith text matn, transmission chains isnad, and claimed sources. Controlled Quranic and Hadith augmentation, together with Shamela-based hard-negative selection, substantially improves citation span detection over training on competition data alone. For Subtask 2, we use source-specific verification pipelines: Quranic evidence is verified through Arabic normalization and approximate matching, while Hadith evidence is verified using BGE-M3 dense retrieval, FAISS candidate search, lexical reranking, and semantic and entity-level evidence. SanadAI achieves 97.35% Macro-F1 on Subtask 1 and 86.66% official macro-accuracy on Subtask 2 on the hidden test set. Quranic verification is strongest at 96.89%, whereas isnad verification remains the most challenging at 77.42%, reflecting the difficulty of verifying transmission chains when the system relies on narrator compatibility with a retrieved Hadith rather than explicitly modeling narrator order and chain continuity.

Islamic question-answering systems can produce fluent Arabic while altering Qur’anic wording, Hadith text, narrator chains, or source attribution. We present Tathabbut for two IslamicEval 2026 settings: fine-grained verification and canonical correction. For verification, Canonical-Provenance Evidence Fusion (CPEF) classifies Ayah, matn, isnad, and claimed-source segments by combining canonical retrieval with collection, narrator, response-level, contradiction, and exact-matn provenance evidence. CPEF achieves 0.9452 macro accuracy on the official Test set. For correction, Type-Aware Selective Canonical Retrieval (TASCR) reranks authoritative Qur’anic and Hadith candidates using lexical, memory-based, contextual, and source evidence, and abstains when confidence is insufficient. TASCR achieves 86.29% Test accuracy. A post-hoc Development diagnostic finds the gold correction for 319/327 (97.55%) recoverable cases in an expanded candidate pool, indicating that candidate discrimination and abstention are larger bottlenecks than retrieval recall. Neither submitted component invokes a generative LLM at inference time.

Mitigating factual hallucinations in Large Language Model (LLM) responses is critical, especially when handling sacred Islamic text. This paper presents our system for the IslamicEval 2026 Shared Task-Subtask 1 (Span Detection), which requires extracting exact character-level spans and fragment labeling of religious citations in Arabic LLM-generated responses. We frame this as a supervised token-classification task, fine-tuning a monolingual bidirectional encoder (AraBERTv2). To process long sequences, we implement a sliding window chunking mechanism with an overlapping stride. We reconstruct final character boundaries using a post-processing heuristic that resolves overlapping token predictions. Furthermore, we apply a cascading false-positive filter using cross-type conflict resolution and category-specific minimum length thresholds. Our pipeline yields highly competitive results, achieving a Macro-F1 score of 0.9652 on the development set and a peak evaluation score of 0.9736 on the official-task test set.

We describe the Qahera system for the four IslamicEval 2026 subtasks: (1) detecting Qur’an/Hadith citation spans in LLM-generated Islamic answers, (2) judging citation authenticity, (3) correcting inauthentic citations against reference corpora, and (4) judging whether an authentic citation answers the question asked. Under a 13B-parameter open-model constraint, running entirely on free Kaggle GPUs, we combine a lightweight token-classification model with character-level auxiliary features (Sub-task 1), corpus-grounded fuzzy retrieval (Sub-tasks 2–3), and a 4B-parameter instruction-tuned LLM (Gemma-3-4B) used only where retrieval alone cannot discriminate (Subtasks 3–4). Rather than hand-designing LLM prompts once, we develop them through an iterative propose-evaluate-diagnose search loop that reads full prediction traces after every candidate and hunts for systematic biases before each change. This surfaces small-model failure modes – a near-universal “reject” bias in isolated yes/no verification, and an “always name something” bias in numbered-list prompts – that would otherwise silently cap performance, and shows that validating candidates on small subsets can be actively misleading. Our final system reaches 0.9716 F1 (Subtask 1), 0.906 accuracy (Subtask 2), 0.773 accuracy (Subtask 3), and 0.673 macro-F1 (Subtask 4) on the official test sets.

Reliable use of large language models in religious settings requires more than fluent text, since a generated answer may misquote Qur’anic verses or Hadiths, assign them to the wrong source, or provide an inaccurate narrator chain. This paper presents the Taqwatech systems for Subtasks 1 and 2 of IslamicEval 2026. For Subtask 1, we fine-tune CAMeLBERT-Mix as a BIO sequence-labeling model to detect Ayah, Hadith matn, isnad, and claimed-source spans. The predicted token labels are then converted into the character offsets required by the task. For Subtask 2, we develop a corpus-grounded verification system that combines Arabic normalization, lexical retrieval, character-level TF–IDF, Qur’anic verse matching, source validation, and narrator matching. On the official test set, our systems achieved a character-level macro-F1 score of 0.9683 for Subtask 1 and a macro accuracy of 0.9050 for Subtask 2. The results show strong performance in span detection and Qur’anic citation verification, while isnad verification remains the main area for improvement.

Large language models may generate inaccurate Islamic citations when answering religious questions. We present the Namaa Community system for Subtask 2 of IslamicEval 2026, which identifies hallucinations in Qur’anic verses (Ayah), hadith texts (matn), chains of narration (isnad), and claimed sources. We formulate the task as retrieval-grounded verification: citation spans are normalized and matched against the Qur’an and the six major hadith collections using character n-gram retrieval and edit-distance re-ranking. Type-specific verifiers assess Ayah and matn similarity, while claimed sources and isnads are checked against their parent citations. The deterministic pipeline requires no language model or GPU, and every decision is traceable to a canonical record. It achieves a macro accuracy of 0.846 on the development set, 17.9 percentage points above a per-type majority-class baseline, and 0.668 on the official blind test; we trace most of that shortfall to a submission-time configuration error and decompose the remainder by segment type. Our code and submissions are publicly available at https://github.com/NAMAA-ORG/Namaa-Community-at-IslamicEval-2026.

We present HalluScoring 2026, a shared task for evaluating hallucination detection and factual verification in Arabic question answering under challenging generalization settings. The shared task is organized into two main tasks, each comprising two subtasks, for a total of four subtasks. Task 1 evaluates binary hallucination detection, considering generalization to unseen questions (Subtask 1.1) and responses generated by unseen LLMs (Subtask 1.2). Task 2 extends the evaluation beyond detection by requiring the systems to additionally identify the correct factual answer from six related candidates, covering Islamic knowledge (Subtask 2.1) and general knowledge (Subtask 2.2). The shared task is based on two Arabic datasets: HalluScore and HalluTruthQA. A total of 13 teams participated in the shared task, 10 of which submitted system description papers. The results of Task 1 demonstrate that hallucination detection remains challenging under distribution shift, with the winning team achieving AUC-ROC test scores of 0.772 and 0.767 for Subtasks 1.1 and 1.2, respectively. For Task 2, the winning team achieved scores of 0.882 and 0.857 in the Islamic and general-knowledge subtasks, respectively, under assisted evaluation.

We describe our submission to HalluScoring 2026 for Arabic hallucination detection in general-culture and Islamic-knowledge question answering. Our system adapts the instruction-tuned 12B-parameter Gemma 4 model with QLoRA. Its central idea is to use the Hashtag AI knowledge-graph retrieval system only during offline data construction: we query facts related to development-set failures, create contrastive training examples, and distill them into the model parameters. The submitted model is closed-book and uses no retrieval, search, or external knowledge at evaluation time. On organizer-provided test annotations, it obtains 73.59 macro-F1 on the 400-example Islamic set and 77.56 on the 1,600-example mixed-domain set; answer-option accuracies are 75.25% and 81.13%. Analysis shows strong science performance but lower recall for subtle hallucinations in history, geography, and Islamic knowledge.

HalluScoring 2026 Subtask 2.2 (Hallucination Detection & Find the Truth) requires systems to judge whether an Arabic generated answer is hallucinated and to select the correct one among six candidates, without access to the reference answer that would make both judgments trivial. We treat that withheld reference as privileged information: a teacher is fine-tuned with it in the prompt, then never deployed, and its advantage is transferred to a blind student through two channels knowledge distillation over the teacher’s output distribution, and a contrastive privileged-information (CPI) term that aligns the student’s hidden representation with the teacher’s gold-conditioned one. Our fine-tuned ALLaM-7B attains 84.62 combined score (Comb) on the official test set and 83.47 on the dev set.

We present our submission to HalluScoring 2026, an ArabicNLP shared task on hallucination detection and answer recovery in Arabic LLM outputs. For Track 1, our submitted baseline fine-tunes CAMeLBERT for binary hallucination detection, achieving 0.921 and 0.763 AUC-ROC on the development and test sets of Task 1.1, and 0.749 and 0.729 on Task 1.2, respectively. We further conduct post-submission ablations with intermediate-layer supervision and category embeddings to study their effect on generalization. The combined model achieves 0.929 AUC-ROC on Task 1.1 and 0.777 on Task 1.2 development sets. For Track 2, we separate hallucination detection from answer selection, using CAMeLBERT for detection and a decoder-based, QLoRA-tuned AceGPT-v2-8B-Chat for option selection. The system achieves combined scores of 0.784 on Task 2.1 and 0.829 on Task 2.2 on the validation sets, improving substantially over the corresponding zero-shot baselines.

Reliable Arabic hallucination detection requires not only strong ranking performance but also robustness to distribution shift, calibrated probabilities, and useful uncertainty for selective prediction. We present the NAMAA Community’s submission to HalluScoring 2026 Task ~1.1 and a set of post-submission diagnostic experiments. The scored, template-compliant system is a CAMeLBERT cross-encoder over the question and generated answer; it achieves 0.7596 ROC-AUC on the hidden test set, ranking third among seven teams. We compare this template-defined input path with reference-aware answer-pair models that use the gold answer. A CAMeLBERT input-framing ablation raises the unseen-question ROC-AUC from 0.8713 to 0.9272, while similar gains appear across ARBERT and MARBERTv2. Reference-free distillation recovers only a small portion of this advantage, whereas LoRA-InfoNCE substantially improves direct cosine discrimination. On a separately trained CAMeLBERT reliability replicate, temperature scaling reduces expected calibration error from 0.0993 to 0.0562, and MC-Dropout variance reduces retained-set error from 14.08% to 8.35% when the 30% most uncertain cases are deferred. These results motivate reporting distributional generalization, calibration, and selective risk alongside ROC-AUC. Code and checkpoints for the scored submission are publicly available. [<https://github.com/NAMAA-ORG/HalluScoring-Arabic-Hallucination-Detection>]

This paper presents a multistage ensemble-based approach developed for the HalluScoring-2026 shared task on Arabic hallucination detection. The proposed system combines three Arabic pretrained language encoders: MARBERTv2, CAMeLBERT, and ARBERT. The diverse linguistic representations captured by these models enable the ensemble to improve the robustness and discriminative capability of hallucination detection compared to the individual encoders. On the development set, the ensemble achieved an AUC-ROC of 0.9413, Macro-F1 of 0.8704, and an AUC-PR of 0.9610, outperforming the individual models. On the official test set, the system achieved an AUC-ROC of 0.7717, Macro-F1 of 0.6576, and an AUC-PR of 0.9181. These results secured the highest overall ranking among all submitted systems, placing our system first in the HalluScoring-2026 shared task. Further analysis reveals substantial variation in detection performance across generator models and linguistic phenomena, with subtle factual deviations, dialectal and culturally specific expressions, and unusually long responses posing particular challenges. Overall, the findings demonstrate the effectiveness of combining distinct Arabic-specific pretrained encoders for robust Arabic hallucination detection.

Hallucination detection in Arabic Islamic question answering is challenging because fluent answers may misstate a verse, narration, ruling, attribution, or historical fact. The task jointly requires detecting hallucination and recovering the factually correct option. We introduce C3R, a Counterfactual Cross-Check Recognition framework that models this relation with two role-conditioned predictions from one Arabic language model: a truth role selects the factual option, while a match role selects the candidate that best represents the generated response. Training uses targeted counterfactual responses, hard distractors, option-order permutations, and completion-likelihood supervision without a cross-view consistency loss. At inference, C3R scores valid structured completions, maps predictions from permuted presentations back to semantic option identities, and derives the hallucination label from argmax agreement between the two roles. The cross-check is applied only at inference. With a QLoRA-adapted ALLaM backbone, C3R ranks first on the official test leaderboard with a final score of 0.8824, comprising 0.8673 macro-F1 and 0.8975 option accuracy. These results demonstrate the effectiveness of structured cross-role agreement for knowledge-sensitive hallucination detection and answer verification.

Large language models (LLMs) produce fluent but factually incorrect outputs which become significant when such models are applied to local languages, such as Arabic, because of the language’s complex morphology, dialectal richness, and the insufficient material available for training and evaluation. The competition HalluScoring-2026 Track 1 aims to address these problems and consists of two tasks that demand hallucination detection in Arabic question-answering settings. To solve this issue, we transfer the knowledge of a powerful reference-based teacher that is a stacked ensemble of five fine-tuned encoder classifiers with dynamic lexical, embedding, and entailment characteristics onto a simpler reference-based student model.

The rapid advancement of large language models (LLMs) has enabled powerful AI applications, yet hallucinations remain a major challenge to their reliability. As part of Track 2 of the HalluScoring 2026 shared task at ArabicNLP 2026, we present a system for detecting hallucinated answers and selecting the correct answer from six closely related candidates. We adapt the Arabic–English ALLaM-7B-Instruct model using Low-Rank Adaptation (LoRA) with domain-specific supervision from Qur’anic text, Hadith, Islamic law and inheritance, Islamic history, and general knowledge. The system jointly performs hallucination detection and answer selection, achieving 88.9% accuracy and 88.6% macro-F1 for hallucination detection, and 91.1% answer selection accuracy on the 800-instance development set. Error analysis identifies geography and structured-output parsing as the main remaining weaknesses.

Large language models answer fluently but not always truthfully, and catching a wrong answer is hardest in knowledge-sensitive domains such as Islamic question answering, where a confi dent but fabricated citation of Qur’an, Hadith, or jurisprudence is easy to produce and costly to miss. We study this through our participation in the HalluScoring 2026 shared task, Task 2 (Islamic Knowledge) (Alansari et al., 2026): given a question, a baseline-generated answer, and six candidate options, a system must de cide whether the answer is hallucinated and then select the correct option. Our system is a single QLoRA adapter fine-tuned on Fanar-1 9B to reproduce the organizers’ two-step infer ence, under a strictly leak-free setup, where no answer information beyond the question and generated answer enters the detection prompt, following our hypothesis that answer exposure inflates detection scores through leakage rather than judgment. On the official leaderboard our submission placed second of five teams, scor ing 0.7582 on the blind track we targeted and 0.7881 on the assisted track. Controlled follow up experiments support it: showing the options during detection is worth +0.047 Macro-F1 on both splits, and detection ability comes entirely from supervision, the frozen generator accept ing its own output on all 400 test items. A 1,200-item probe, however, corrects our own initial diagnosis. The model is not blind to clean falsehoods, catching 80.5% of them; it is biased toward rejection, wrongly flagging 32.5% of verified-correct answers, and that sin gle bias is the central limitation of the approach.

Large language models used for question answering can produce fluent but factually incorrect answers, a problem that is acute in knowledge-sensitive domains such as Islamic religious knowledge. We describe our system for the HalluScoring 2026 shared task (Islamic Knowledge track), which requires jointly detecting hallucinated answers and selecting the correct option among six candidates. We fine-tune Fanar-1-9B-Instruct with LoRA and decompose inference into two independently-prompted steps, judging hallucination before revealing the candidate options, to prevent a textual-similarity shortcut where one option is a near-verbatim restatement of the generated answer. Our system scores 0.7149 (Assisted) and 0.6222 (Blind) on the official hidden test set, ranking 5th of 5 reported teams; a controlled comparison shows this shortcut-resistant design costs roughly 10 points against a single-pass variant on a held-out development subsample (0.7625 vs. 0.8639 Combined Score). We release our code and model to support further work on this trade-off.

We present KnowledgeGraphEval 2026, the first Arabic shared task dedicated to knowledge graph (KG) construction from text. The task focuses on extracting named entities and semantic relations from Arabic text to build structured, machine-readable knowledge representations. It includes two subtasks: (i) AdaptNER, which evaluates cross-domain named entity recognition (NER), and (ii) Relation Extraction (RE), which predicts semantic relations between predefined entity pairs. For AdaptNER, we release KONOOZ MSA, an Arabic NER benchmark annotated with 21 entity types across 10 domains. For RE, we release WojoodRelations, an Arabic RE dataset annotated with 40 relation types. At test time, 14 teams participated, of which 13 submitted system description papers, and one RE paper was subsequently rejected. The best systems achieve F1 scores of 76% and 96% on AdaptNER and RE, respectively. Results show that cross-domain NER remains challenging, while RE systems continue to face difficulties with rare and semantically confusable relations. We release the datasets, baselines, and evaluation tools for reproducible research in Arabic information extraction and KG construction.

This paper presents the submission of team AdaptLinker to the AdaptNER shared task (Subtask 1: Cross-Domain Named Entity Recognition) at KnowledgeGraphEval 2026. To address nested entities and domain shift across 10 unseen target domains, we propose a multi-head token classification framework built on the AraBERT backbone. Our approach inte- grates: (1) 21 parallel linear classifiers to pre- dict overlapping named entities; (2) a Domain Adversarial Neural Network (DANN) using a Gradient Reversal Layer (GRL) to align feature spaces; (3) confidence-gated pseudo-labeling on unlabeled target text; and (4) Stochastic Weight Averaging (SWA). Our final system achieved a Precision of 75.48%, Recall of 71.21%, a Micro F1 of 73.29%, and a Macro F1 of 63.37% on the official blinded test set, outperforming the shared-task baseline (0.7329 vs. 0.622 Micro F1) and ranking 6th out of all participating teams

Arabic knowledge graph construction is a challenging problem due to the linguistic complexity of Arabic and the need for robust generalization across domains. This paper presents our systems for the KnowledgeGraphEval 2026 shared task on Arabic knowledge graph construction and domain adaptation, covering two subtasks: AdaptNER for cross-domain named entity recognition and relation extraction (RE). For AdaptNER, we employ an entity-span ensemble combining two BIO sequence taggers with constrained conditional random field (CRF) decoding and a GlobalPointer span classifier. Our system achieved a micro-F1 score of 76.00%, ranking first in the official evaluation. For relation extraction, we use a majority-vote ensemble of three LoRA-fine-tuned causal language models. This approach achieved a micro-F1 score of 93.48%, ranking fifth. These results demonstrate the effectiveness of model ensembling and complementary prediction strategies for Arabic named-entity recognition and relation extraction.

This paper presents the PTUK-NLP submission to KnowledgeGraphEval 2026 Subtask 1 (AdaptNER), evaluating cross-domain and cross-dialect nested Arabic Named Entity Recognition (NER). Models trained on Wojood are evaluated in a zero-shot setting on Konooz, spanning ten unseen domains and sixteen Arabic dialects. Our system fine-tunes ArBERT with targeted augmentation for rare entity types and lightweight inference-time post-processing, using Konooz strictly for evaluation and never for training or adaptation. The submitted system achieves 91.07% micro-F1 and 83.11% macro-F1 on Wojood, and 71.84% precision, 66.59% recall, 69.11% micro-F1, and 60.70% macro-F1 on the official Konooz blind test, ranking fourth on the final leaderboard and achieving a micro-F1 score 6.11 points above the organizers’ baseline. A post-submission ensemble of ArBERT and AraBERTv2 does not improve over the best single model. Overall, targeted data-centric modifications provide useful gains for rare entity types, while a standard MSA-pretrained encoder remains competitive under substantial domain and dialect shift.

Relation extraction (RE) is a fundamental component of knowledge graph construction, question answering, and text summarization. Although recent RE models achieve strong overall performance through the fine-tuning of pretrained language models, their aggregate performance can obscure substantial disparities across relation types, particularly under long-tailed distributions. This issue remains largely unexplored for Arabic RE. In this work, we investigate the effectiveness of class-balancing strategies for long-tail relation extraction in Arabic by fine-tuning AraBERT with entity markers on Wojood^Relations. Our system, submitted to the KnowledgeGraphEval 2026 shared task, achieves a Micro-F1 score of 92.73 on the official test set. Further analysis shows that errors are disproportionately concentrated among low-frequency and semantically overlapping relation types, indicating that imbalance-handling strategies developed primarily for English benchmarks do not fully address the challenges posed by the long-tailed relation distribution in Arabic.

We describe Team Wasl’s systems for the KnowledgeGraphEval 2026 shared task at ArabicNLP 2026, which decomposes Arabic knowledge-graph construction into two evaluated stages: nested NER under domain shift and relation extraction between given mentions. For Subtask 1 (AdaptNER), we train a multi-head BIO tagger (one three-way O/B/I head per entity type over a shared pretrained Arabic encoder) on Wojood and combine three random seeds by span-level majority vote. For Subtask 2, forced-choice classification over the 40 WojoodRelations relation types, we fine-tune a typed-marker classifier with type-constraint decoding, ensembling five grouped folds by logit averaging. Our submissions scored 0.7483 micro-F1 on AdaptNER’s blind Konooz slice (ten unseen domains), 4th of the eight officially ranked teams and 11.8 points above the official baseline, and 0.9571 micro-F1 on relation extraction with the no-relation class masked at decoding, 2nd at the results release; re-scored by the organizers over all 41 classes, the same ensemble reaches 0.9328, 4th in the organizers’ final ranking. We analyze where the 18.6-point domain-shift tax on NER falls: chiefly on recall. We further show that type-constraint decoding and ensembling are partially redundant (+0.9 points on a single RE model, +0.1 on the ensemble) and that the surviving errors share an argument-type signature that no type constraint can separate.

Automated knowledge graph (KG) construction involves two core components: nested named entity recognition (NER) and relation extraction (RE). In this work, we address cross-domain nested NER alongside RE. We integrate domain adaptation methods within a mixture-of-experts ensemble system for nested NER from Arabic text, evaluated in a cross-domain setting. The system combines multiple Arabic pretrained encoders for token- and span-level classification with domain data augmentation, and integrates their predictions using a vote-based ensemble routing with a complementary XLM-R expert. For RE, we employ an adversarially trained relation classifier backed by auxiliary entity types, NLI reranking, and structural schema rules, treating no_relation as a standard relation label. We additionally study a positive-relation setting that excludes no_relation from both training and inference. Our NER system reaches 0.7438 micro F1 on the hidden test set. For RE, the positive-relation setting achieves 0.9608 micro F1, while the no_relation-aware system achieves 0.9295.

Relation extraction (RE) identifies semantic relations between entities, converting unstructured text into structured facts for downstream applications such as knowledge graph construction. Arabic RE remains challenging due to rich morphology, complex syntax, dialectal variation, and the scarcity of high-quality annotated resources. We present Yafa’s Team submission (ranked 7th) to Subtask 2 of the KnowledgeGraphEval 2026 Shared Task, which frames Arabic RE as multiclass classification over 40 relation types from the dataset. We propose an ArBERTv2-based entity-marker classifier that fuses global sentence context ([CLS]) with entity-specific representations (h_e1, h_e2) into a joint representation for relation prediction. Building on this baseline, we exploit a transductive self-training step that pseudo-labels the released test inputs at a softmax confidence threshold of 0.95 and retrains on them, raising Micro-F1 from 70.20% to 91.92%. We further conduct a systematic analysis of cross-backbone ensembling with MARBERTv2 and AraBERTv2, finding that ensembling does not outperform the single transductive model. Our results highlight transductive self-training as a simple yet effective strategy for improving Arabic RE robustness.

Cross-domain Arabic Named Entity Recognition (NER) remains challenging because models trained on one domain can lose performance when applied to unseen domains with different vocabulary, writing styles, and entity distributions. This paper addresses this challenge, and presents Yafa’s Team submission (ranked 7^th) to Subtask 1 of the KnowledgeGraphEval 2026 Shared Task. Our system leverages an AraBERTv2-based model for nested Arabic NER to approach it. The proposed approach combines task-adaptive (TAPT) pretraining on unlabeled Konooz data, R-Drop fine-tuning on Wojood, strong data augmentation, teacher-student knowledge distillation, span-level confidence and entropy filtering, and an entity-specific confidence-aware weighted-logit ensemble. Three teacher models generate soft labels for Konooz, and three distilled student models are combined in the final ensemble to produce the final predictions. Wojood contains approximately 448K tokens and 72K entity mentions across 21 entity types, while Konooz contains data from 10 domains for cross-domain evaluation. The final system achieves a Micro F1 of 71.33% on the official Konooz test set, improving over the non-adapted AraBERTv02 baseline of 61.00% by 10.33 points. The results show that combining domain adaptation, filtered knowledge transfer, and entity-specific ensembling can improve cross-domain Arabic NER performance.

The ANLP_UniSo team submitted a system to Subtask 1 (Named Entity Recognition) of KnowledgeGraphEval 2026, which focuses on nested NER in Arabic and evaluates crossdomain and cross-dialect generalization from the Wojood corpus to the multi-domain Konooz benchmark. The system is a combination of an Efficient GlobalPointer decoder with rotary position embeddings and a MARBERT backbone. To reduce the domain gap, first domainadaptive pretraining (DAPT) is performed on unlabeled Konooz text, and then a pseudolabeling self-training stage is applied. A second model based on AraBERT is trained and combined to the other model using a stacking ensemble to improve prediction diversity. As for the Wojood test set, the best ensemble achieves a micro F1 of 0.9288, increasing to 0.9292 when entity-type-specific decision thresholds are used, compared with 0.9211 for the single MARBERT model. On the official blind Konooz test set, however, the submitted system obtains a micro F1 of 0.7106 (macro F1 = 0.6101, weighted F1 = 0.7013, P = 0.8143, R = 0.6304, accuracy = 0.9948). The lower recall suggests that the decoding threshold chosen based on the Wojood validation data was too strict for the domain-shifted Konooz data. the system study the difference and propose a threshold-recalibration strategy to improve the performance under the target-domain distribution.

In this paper, we present a relation extraction system developed for Subtask 2 of KnowledgeGraphEval 2026. The system fine-tunes a single ARBERTv2 encoder on inputs with typed entity markers that expose entity boundaries and types, also classifies each entity pair from the encoder’s classification token ([CLS]) state concatenated with the states at the subject and object start markers. It is trained over the full 41-label space, including no-relation, with focal loss, class-balanced sampling, label smoothing, and no-relation down-weighting to counter severe class imbalance. Our submission ranked third of seven teams with a micro-F1 of 0.9332, 0.23 points above the official baseline, which fine-tunes the same encoder. We also report an analysis of seed-dependent behaviour on fine-grained geopolitical relations from our earlier development experiments and discuss the limitations of our evaluation.

This paper details our submission to KnowledgeGraphEval 2026 Subtask 2 on directed Arabic relation extraction. We explored encoder-based and large language model approaches for predicting one of 40 relations or no_relation from an Arabic sentence and a subject–object pair. Our encoder systems use typed entity markers and separate relation-existence and relation-classification heads, while our LLM systems fine-tune Yehia and Qwen3.5 with QLoRA and ontology-constrained prediction. Our best system, an ensemble of Yehia-RE, Qwen-RE, and TypedPair, achieved a micro-F1 score of 94.06 and ranked 1st in Subtask 2. Yehia-RE (full-vocabulary) achieved 92.98, while the best encoder and Qwen-RE achieved 92.69 and 91.94 respectively. Our results highlight the importance of constrained prediction and complementary modeling approaches for Arabic relation extraction.

This paper presents the ArmyForNER system for Subtask 1 of the KnowledgeGraphEval 2026 shared task on Arabic named entity recognition (NER). We address nested, cross-domain Arabic NER with a multi-stage pipeline built on ARBERTv2: (1) domain-adaptive pre-training (DAPT) via masked language modelling on unlabelled Konooz text, (2) fine-tuning with a multi-head BIO tagging architecture on the Wojood nested NER corpus, and (3) self-training on pseudo-labels produced by an intermediate DAPT ensemble over the Konooz domains. Soft-voting three self-trained checkpoints obtains 74.85 micro-F1 and 65.20 macro-F1 on the official test set, against a 63.00 / 44.00 base- line. Because both adaptation stages consume the blind Konooz test sentences, our system is listed among those that used Konooz during training, and its score is not comparable to systems that used Konooz for evaluation only. We state this setting explicitly and find that self-training gives consistent gains across seeds, whereas the benefit of DAPT is configuration- and seed-dependent.

This paper presents Daleel 2026, the first shared task on argument mining in Arabic news editorials and competitive debates. Daleel includes two tasks: Task 1 identifies which of six argumentative discourse unit (ADU) categories are present in a paragraph, while Task 2 locates the units with their labels and exact text spans. The tasks are performed using a new corpus of 30 documents, 1,042 paragraphs, and about 76k words. Twenty-nine teams registered, 12 participated, and 10 submitted system papers. Both tasks have Closed and Open tracks. The best results for the Closed track were 0.7117 macro-F1 in Task 1 and 0.7547 modified overlap F1 in Task 2. The results show different insights: the best Closed-track score of each team was higher on debates than on editorials. A broad range of approaches was used in Task 1 submissions, including Arabic encoders, large language models, rules, and ensembles. Task 2 systems, in particular the top-ranked ones, were more similar and mainly used CRF or Viterbi decoding with combinations of models or training runs. The rare ADU categories remained difficult in both tasks, and identifying exact span boundaries was an additional challenge in Task 2.

We present BME-Daleel, our system for both Daleel 2026 Arabic argumentative discourse mining tasks. A joint Transformer predicts paragraph-level label presence and token-level BIO states; constrained decoding and domain-routed ensembles map spans to original character coordinates. Repeated out-of-fold comparisons gate domain specialists, while routing selected decoded outputs preserves other-genre rows exactly. Using organizer annotations and public pretrained models, our Closed-track system scores 0.7117 macro-F1 for paragraph classification and 0.7547 modified F1 for span detection.

Argument Mining (AM) is a critical NLP task that remains significantly under-resourced in Arabic. This paper presents STAR-Ar, a BERT-BiLSTM-CRF architecture for argument discourse detection and classification, as our system for Daleel 2026, the inaugural Arabic argument mining shared task. The task requires the identification and classification of argumentative discourse units (ADUs) in debate and editorial texts. We jointly model these two objectives as a token-level sequence labeling task using a BERT-BiLSTM-CRF architecture that combines contextual transformer embeddings with structural transition constraints to support accurate span detection. STAR-Ar achieves an F1-score of 72.69 on validation and 73.7 on test data. Our domain-specific analysis shows that models trained exclusively on editorials underperform those trained on debates, a disparity we primarily attribute to the smaller size of the editorial dataset.

We describe the SalahAbdo closed-track systems for both Daleel 2026 subtasks: deciding which argumentative-unit types appear in an Arabic paragraph (Task 1) and locating their exact spans (Task 2), in editorials and debates. No single method was best for all six argumentative-unit types, so our Task 1 system routes each type to the method that measured best for it: assumption (AS), testimony (TE), and statistics (ST) are detected by sampling a gemma-4-31b classifier five times and firing when enough samples agree; anecdote (AN) and other (OT) by a fine-tuned CAMeLBERT classifier; and the lowest-agreement type, common ground (CO), by ranking paragraphs and firing a fixed count. A final verifier shows the LLM examples of true and false ST paragraphs and removes likely false positives. Task 2 first extracts spans as verbatim LLM quotes, then re-labels them by blending encoder probabilities with label-sequence statistics under Viterbi decoding, with zero additional LLM calls. Every design change had to clear an adoption bar set before its result was read: 29 recorded decisions produced six adoptions, and composition arithmetic over already-scored legs predicted two development scores digit-for-digit. Development scores are 0.7089 macro-F1 (Task 1) and 0.7247 span F1 (Task 2); the pinned recipes scored 0.6499 and 0.7055 on test, and our final official submissions scored 0.712 and 0.7316: first of ten and fifth of eight closed-track teams (Section 6 discloses the evaluation-phase protocol). Residual errors track the scheme’s measured human agreement (kappa as low as 0.114): types whose definitions a model can check against the text improved under LLM verification, while types defined by annotation convention resisted every prompting attack and were learned only from supervision and sequence structure.

We present our system for Daleel 2026 Task 2 on Arabic Argumentative Discourse Unit (ADU) detection. Our main contribution is a Conformer-based tagging architecture that combines a pretrained AraBERT encoder with a Conformer head to capture both global discourse context and local span-boundary cues. To the best of our knowledge, this is the first application of a Conformer-based architecture to Arabic ADU detection. The system also uses overlapping token windows, class-weighted training, character-level span reconstruction, and lightweight post-processing. On the official hidden test set, our system achieved an overall F1 score of 0.732, substantially outperforming the official baseline of 0.285 and ranking fourth in the closed track.

We describe the Durham-Jazan submissions to the Daleel 2026 shared task on Arabic argumentative discourse mining, which comprises multi-label paragraph classification (Task 1) and argumentative span detection (Task 2), each with a Closed and an Open track. Our Task 2 Closed system routes the Modern Standard Arabic editorials to CAMeLBERT and the noisier debate transcripts to MARBERTv2. The Open system blends the two families at character level and adds synthetic training data from a large language model for rare-class generalisation, generated by a segment-concatenation recipe that makes span offsets exact by construction. The Task 1 systems add back-translation, targeted augmentation of the rare classes, and a QLoRA-tuned Qwen3-32B for the editorials. Everything else trains on one 16 GB consumer GPU. Our official test scores are 0.729 and 0.736 overlap-F1 on Task 2 and 0.657 and 0.669 macro-F1 on Task 1 (Closed and Open). We placed first of two on Task 1 Open, first of one on Task 2 Open, fifth of ten on Task 1 Closed, and sixth of eight on Task 2 Closed. Our central finding concerns our own systems on this data. Recalibrating per-class thresholds on validation outputs overfits them. Under a controlled protocol the held-out effect is negative at every tuning-set size we tested, while the apparent in-sample gain shrinks as the tuning set grows. An audit of eighteen interventions from our campaign is consistent with this, though observational. A size-matched study links the editorial deficit that every team showed to the genre composition of the training data. We release our code and synthetic data.

This paper presents our submission to Task 1 (Argumentative Discourse Unit Classification, Closed Track) of the Daleel 2026 Shared Task. The task addresses Arabic multi-label argumentative discourse unit classification at the paragraph level. Our system fine-tunes CAMeLBERT-Mix using a class-weighted Binary Cross-Entropy loss, paragraph-type-aware input representation, head-tail tokenization for long paragraphs, and label-specific threshold optimization. On the official development set, the proposed system achieves a Macro F1-score of 0.6121, a Micro F1-score of 0.7509, and a Samples F1-score of 0.6387. These results suggest that a carefully optimized CAMeLBERT-Mix framework provides a competitive and reproducible approach for Arabic argumentative discourse unit classification under the Closed Track setting.

This paper presents our systems for the Daleel 2026 Shared Task. We explore multitask learning, domain-specific ensembles, and a two-stage BIO segmentation and classification pipeline with LLM-based disambiguation, alongside classical machine-learning, Transformer, and LLM baselines. Results show that supervised Transformers outperform standalone LLM prompting, while LLMs are more effective for selectively resolving ambiguous predictions. The multitask model generalizes best on Task 1, achieving 0.599 F1, while the ensemble multitask model reaches 0.683 F1 on Task 2. Our results show the potential of LLM integration within Transformer-based systems for Arabic argumentative discourse mining.

We present our closed-track systems for both subtasks of Daleel 2026, which targets Arabic argumentative discourse unit (ADU) mining over editorials and competitive debates. Task 1 is an imbalanced multi label paragraph classification problem, and Task 2 requires character-level detection and classification of ADU spans. For Task 1 we investigate two parameter-efficient decoder-LLM formulations: joint label generation with Qwen3.5-9B and independent label scoring with ALLaM-7B. The latter decomposes each paragraph into six explicit Yes/No decisions, giving every label direct supervision and a separate decision score; it achieves 0.6714 Macro-F1 on the evaluation phase and ranked 4th, while ablations show that additional class-balanced weighting degrades performance. For Task 2 we combine task-adaptive pretraining, Beginning-Inside-Outside Conditional Random Field (BIO-CRF) sequence labeling, boundary-jitter augmentation, character-space ensembling, and span post-processing over AraBERT-large, achieving an official F1 of 0.7523 (0.7931 on debates, 0.6417 on editorials) and ranking second. Across both tasks, our results support explicit label-wise supervision and span-level ensemble decisions as effective ways to handle label imbalance and boundary noise without aggressive reweighting, while short spans, rare classes, and genre variation remain challenging.

This paper presents our system for Task 1 of the Daleel 2026 shared task, which assigns multiple discourse-function labels to Arabic paragraphs from editorials and debate transcripts. We fine-tune a monolingual Arabic encoder (MARBERTv2) with a label-weighted focal loss, multilabel-stratified k-fold crossvalidation, out-of-fold per-label threshold calibration, k-fold ensembling, and a genreconditioning prefix, in a two-stage protocol that first trains on the training split alone and then retrains on training+development. Internal cross-validation macro-F1 rises from 0.667 to 0.680 across the two stages. On the official test set, the system attains micro-F1 = 0.766 and macroF1 = 0.560, with markedly stronger micro-F1 on debate paragraphs (0.825) than editorials (0.633), and a smaller but consistent macro-F1 gap (0.554 vs. 0.456). Assumption is classified most reliably (F1 = 0.88–0.94) and Common Ground least reliably (F1 ≤ 0.12), mirroring the low inter-annotator agreement reported for the same category in the source English scheme. We find a ∼0.12 gap between internal cross-validation and official test macro-F1, concentrated in low-support, genre-sensitive categories. Our code is publicly available at https://github.com/EidAhmed1/daleel2026_task1_project.

We describe our submission to Task 1 (Argumentative Discourse Unit Classification) of the Daleel 2026 shared task on Arabic argument mining, in the closed track. Given a paragraph from an Arabic editorial or debate transcript, the task is to predict the multi-label set of argumentative functions it contains, drawn from six classes: common ground, assumption, testimony, statistics, anecdote, and other. Our system fine-tunes a pretrained Arabic encoder, comparing AraBERT, MARBERT, and CAMeLBERT-mix, augmented with a small side-channel of hand-crafted lexical features motivated directly by the task’s class definitions (e.g., reporting verbs for testimony, digit/percentage patterns for statistics), trained with a class-balanced focal loss and combined into a 5-fold cross-validation ensemble; CAMeLBERT-mix was selected as the final encoder for all submitted configurations. In the dev phase, our Both-domain configuration achieved an overall macro-F1 of 0.5600 (editorial 0.4350, debate 0.4900). On the official test-phase Codabench leaderboard, our Both-domain configuration achieved an overall macro-F1 of 0.5265 (editorial 0.3838, debate 0.4943), ranking 10th; our Editorial-only configuration reached an overall macro-F1 of 0.4544 (editorial 0.3249, debate 0.4509). We report two findings grounded in these results: overall (pooled) macro-F1 consistently exceeds the macro-F1 computed on either genre subset alone, and our editorial-only model unexpectedly scores higher macro-F1 on the out-of-domain debate subset than on its own in-domain editorial subset. We discuss both as likely artifacts of per-class support sparsity at the smaller genre-subset scale rather than genuine cross-domain superiority, though we did not conduct a formal ablation study or systematic error analysis to confirm this, which we flag as the primary limitation of this report. Our code is available at https://github.com/turab-usmani/daleel-task1-closedtrack.

Sentiment analysis is a fundamental problem in Natural Language Processing (NLP). Standard sentiment classification for the Arabic language remains challenging due to the high volume of dialectal Arabic. To advance research in this area, this paper proposes the Shared Task on Sentiment Analysis and Swapping in Arabic Dialects (DialectSentEval), hosted with the Arabic Natural Language Processing Conference (ArabicNLP 2026). This shared task consists of two subtasks: Subtask 1 focuses on multi-class and multi-dialect sentiment analysis, requiring models to identify sentiment polarity across various Arabic dialects. Subtask 2 introduces a generative task for Arabic sentiment swap, challenging models to invert sentiment polarity while preserving core semantics. In this overview paper, we present the motivation, dataset creation, and summarize the main findings from participating models.

Sentiment swap asks a system to rewrite a sentence so that its polarity is inverted while everything else stays the same, two halves of an instruction that pull against each other. This paper describes how the Sabaa system navigates that trade-off for Arabic dialects in Subtask 2 of DialectSentEval 2026. Rather than asking one model for a single correct rewrite, we ask two fine-tuned Arabic sequence-to-sequence models to each propose several, then choose among the pooled proposals with a scoring function built from the signals the task grades us on. The system scores 49.07 BLEU, 69.33 chrF, and 0.7447 sentiment accuracy on the official test set. That last figure is the most informative result here: the system places 90.37% of its outputs at the polarity it was aiming for, and the shortfall traces to inferring each sentence’s original polarity, which the test data withholds. Measured against gold labels on development, that inference is correct only 74.07% of the time, and sixteen points of sentiment accuracy separate our submission from a perfect estimator.

We describe the AKARROUCH submission to DialectSentEval 2026 Subtask 2, which requires reversing the sentiment of an Arabic dialect sentence while preserving its content. We fine-tune AraT5v2 for direction-controlled generation and use it to produce multiple candidates. We then rerank these candidates using CAMeLBERT target-sentiment probabilities and source-to-candidate chrF. The system ranked second in the official evaluation, achieving 78.51% sentiment accuracy, 49.82 BLEU, and 70.12 chrF.

We describe our participation in Subtask 2 of the DialectSentEval 2026 Shared Task, whose goal is to rewrite an Arabic dialectal sentence so its sentiment polarity is reversed while its meaning, entities, and dialect are preserved. Our system, SWAPPER (Sentiment-SWAPping with Parameter-Efficient adapteRs), is an ensemble of LoRA-adapted mBART-50 checkpoints in which the adapters are extended beyond attention into the feed-forward sub-layers, where the actual lexical rewriting is learned. Three design choices carried most of the weight: (i) a minimal Arabic instruction prefix that reserves encoder capacity for content; (ii) feed-forward LoRA; (iii) forcing the decoder’s beginning-of-sequence token to ar_AR so mBART-50 does not drift into other scripts. On the official blind test SWAPPER reaches Sentiment-Style Accuracy 0.6732, BLEU 41.38, and chrF 65.84. As a secondary contribution, we introduce the J-Gap (𝛥 J = J_tgt - J_src), a Jaccard-based copy-collapse diagnostic that cleanly separates genuine sentiment swappers from models that merely copy their input, and explains why superficially strong backbones such as AraT5v2 fail the task despite high surface metrics.

Arabic dialect sentiment swapping requires systems to reverse the affective orientation of an utterance while preserving its core meaning, dialectal form, and fluency. This task is particularly challenging because sentiment may be conveyed through dialect-specific vocabulary, negation patterns, and spelling variation. We present LahjaSensei, our system for Subtask 2 of the DialectSentEval shared task. We investigate two complementary approaches: prompt-based generation with GPT-5.6 and parameter-efficient fine-tuning of Qwen2.5-7B. We evaluate GPT-5.6 under zero-shot, retrieved few-shot, stronger Arabic instructions, and direction-aware configurations. The direction-aware GPT-5.6 pipeline explicitly predicts source polarity before retrieving direction-matched demonstrations and generating the rewrite. For Qwen2.5-7B, we compare pooled QLoRA with linear and TIES merges of direction-specific adapters. Direction-aware GPT-5.6 prompting achieves the highest test sentiment accuracy and remains stable across development and test. Pooled QLoRA produces the highest development BLEU and chrF, while linear merging obtains the highest test BLEU and chrF but lower sentiment accuracy. These findings highlight the central trade-off between sentiment control and preservation of semantic content and demonstrate the value of explicit direction modeling for Arabic dialect sentiment swapping.

This paper describes our submission to Subtask 1 of DialectSentEval 2026, which focuses on three class sentiment classification for Arabic dialectal text. The task aims to classify sentences written in five Arabic dialects into positive, negative, or neutral sentiment categories. We investigate the effectiveness of transformer-based Arabic language models through a series of controlled experiments. Specifically, we compare MARBERTv2 and AraBERT while maintaining identical pre processing, training, and evaluation settings to ensure a fair comparison. In addition, we examine whether explicitly providing dialect information as part of the input improves classification performance.

Experimental results on an internal validation split show that MARBERTv2 consistently outperforms AraBERT on this task. Furthermore, removing the explicit dialect tag from the input slightly improves performance, suggesting that MARBERTv2 can effectively capture dialect-specific linguistic patterns directly from the sentence. Our best configuration, which uses MARBERTv2 with sentence-only input, achieves a Macro F1-score of 0.9394 and an accuracy of 94.24%. These findings demonstrate the effectiveness of dialect-aware pre-trained language models for Arabic sentiment analysis and establish a strong baseline for the official shared task evaluation.

This paper presents the Gradient Descenders system for DialectSentEval Subtask 1 on multi-dialect Arabic sentiment analysis. Because general-purpose models often miss dialectal nuances, we compare three pretrained encoders with distinct linguistic orientations: MARBERTv2, DarijaBERT, and AraBERT. We formulate the task as three-class classification, employing a parameter-efficient head-only adaptation strategy that freezes encoder parameters. Our results show MARBERTv2 outperforms the others, achieving a Macro F1 of 80.11% and Accuracy of 81.90%, securing fifth place overall. These findings emphasize that backbone selection is critical; specifically, broad dialect-oriented pretraining yields the most transferable representations for multi-dialect generalization.

Sentiment analysis of regional Arabic dialects presents a formidable challenge due to severe orthographic variation, morphological complexity, and a scarcity of annotated resources. To tackle these linguistic hurdles within Subtask-1 of the DialectSentEval 2026 Shared Task, team Axiom proposes a heavily regularized, dialect-aware ensemble framework combining two domain-complementary language models: MARBERTv2 and AraBERTv2. Our pipeline incorporates preprocessed dialect tags appended directly to input sequences, Stochastic Weight Averaging (SWA) for flat-minima convergence, and simple probability averaging across model outputs. Empirical results on the hidden test set demonstrate the robustness of this approach, with our system securing 2nd place overall and achieving an official Macro F1-Score of 0.8211.

We present our system for Subtask 1 (three-class sentiment classification) of the DialectSentEval 2026 shared task, which evaluates five Arabic dialects using the Multi-Dialect-Sent (MDS-3) hotel-review corpus. Our approach fine-tunes MARBERT with a stratified train/validation split and class-weighted cross-entropy loss. On the official test set, our system achieves an F1 score of 0.8175 and an accuracy of 0.8419, ranking 3rd on the official leaderboard. We identify a substantial dialect mismatch in the dataset: the training set contains no Lebanese reviews, while approximately 20% of the test set is Lebanese. Adding additional Lebanese reviews from OCLAR partially reduces this gap, but does not eliminate it. The error analysis also shows that misclassifications are concentrated in the neutral class and in reviews that express mixed feelings, often with high confidence. These findings highlight the challenges of cross-dialect sentiment classification under unseen-dialect conditions.

Arabic dialect sentiment analysis is challenging because regional varieties show considerable lexical, morphological, syntactical, and orthographic variation. This paper presents a transformer-based system developed for DialectSentEval 2026 Subtask 1, a three-class sentiment classification task covering Moroccan, Egyptian, Jordanian, and Saudi Arabic. The system independently fine-tunes three pretrained Arabic language models: AraBERTv2, MARBERTv2, and CAMeLBERT, using a common training protocol. The implementation uses an 80/20 stratified train-validation split, a maximum sequence length of 128, a batch size of 16, a learning rate of 2 × 10^-5, and focal loss with 𝛾=2=2. Model selection was conducted based on validation macro F1. Using the published labeled test set, MARBERTv2 achieves the strongest independent result with 94.07% macro F1, followed by CAMeLBERT at 81.96% and AraBERTv2 at 79.77%. The results demonstrate the importance of Arabic pretraining and dialect-oriented representations and establish MARBERTv2 as the strongest standalone backbone in the current system.

Arabic sentiment swap aims to reverse the polarity of a dialectal sentence while preserving its underlying meaning. We present NAMAA’s system for DialectSentEval 2026 Subtask 2. Fine-tuning AraT5v2 achieves 79.56 development sentiment-style accuracy, while reranked sampling over eight candidates, using sentiment agreement and content-overlap signals, raises this to 93.57. Cross-backbone ensembling further reaches a development score of 96.78. On the hidden test set, however, this ordering does not transfer: a widened rerank pool with a re-tuned content weight achieves our best official score of 0.7597, whereas the most complex all-source ensemble scores 0.7343. We identify two likely contributors to this development-to-test mismatch: the absence of gold source-polarity labels at test time, which requires classifier-based polarity inference, and train–development sentence duplication that may inflate development performance. These results show that additional generation and ensemble complexity can improve development metrics substantially without guaranteeing better generalization under shifted evaluation conditions.

We describe our submission to DialectSentEval 2026 Subtask 1, a three-class sentiment task covering Moroccan, Egyptian, Lebanese, Jordanian, and Saudi Arabic. Our system fine-tunes MARBERTv2 in five stratified folds and averages their test probabilities. Using raw text without explicit dialect conditioning or external task-specific data, the official submission obtained 0.7939 macro-F1, 0.8152 accuracy, 0.7917 macro precision, and 0.7987 macro recall, ranking eighth among twelve systems in the supplied snapshot. Separately, a post-submission character 3-5-gram diagnostic falls from 0.9157 OOF macro-F1 with row-stratified folds to 0.7520 when repeated ID-label groups and normalized duplicates are kept together. This result establishes dataset-level split sensitivity; it does not quantify the effect on MARBERTv2. We document the submitted configuration, a group-aware diagnostic protocol, and the experiments still needed to isolate the effects of strict splitting and ensembling on the transformer system.

Arabic sentiment analysis must cope with dialects that differ in vocabulary, morphology and spelling. Subtask 1 of the DialectSentEval 2026 shared task poses this problem directly: systems label hotel review sentences from five Arabic dialects as positive, negative or neutral, and one dialect (Lebanese) appears only in the test set. In this paper, we describe the Lahja submission. Since the training corpus is small (1,731 sentences) and consists of parallel translations, we treat unreliable measurement (i.e. leaky validation splits and single training runs) rather than model capacity as the main risk. We fine-tune MARBERTv2 under three random seeds, average their softmax outputs, and validate on a split grouped by review identifier so that all dialect versions of a review stay on the same side. The ensemble reached 0.790 macro-F1 on the official test set, ranking 10th of 12 teams, and lost only about two points from validation to test despite the unseen dialect. Our ablations show that the variation between seeds (about 1.5 points) exceeds the effect of orthographic normalisation or class weighting, and that the normalised ensemble, although best on validation, generalised worst on test. We therefore recommend grouped validation and seed-spread reporting as standard practice for small-data shared tasks.

We present NU-SentEval’s system for DialectSentEval 2026 Subtask 2 (Arabic Dialect Sentiment Swap), which rewrites Arabic dialectal sentences into the opposite sentiment polarity while preserving their core content. Our approach fine-tunes AraT5v2-base-1024, a pretrained Arabic sequence-to-sequence model, directly on the shared task’s source-target sentence pairs, and generates outputs using standard beam search decoding. On the official Final Evaluation test leaderboard, our system achieves 54.63 BLEU, 72.61 chrF, and 68.98% sentiment preservation. Our manual error analysis shows the model reliably inverts sentiment for short, single-clause inputs, often matching the human reference exactly, though it occasionally fails to propagate the sentiment flip to secondary cues such as emojis, resulting in mixed-signal outputs.

This paper presents the system developed by the NU-SentEval team for Subtask 1 (Sentiment Analysis in Dialectal Arabic) of the DialectSentEval 2026 Shared Task. Arabic dialect sentiment classification is challenging due to morphological complexity, orthographic variation across regional dialects, and class imbalances in user reviews. To address these challenges, we propose an ensembled architecture leveraging five diverse Pre-trained Language Models (PLMs): MARBERTv2, MARBERT, CAMeLBERT-DA, AraBERTv0.2-Twitter, and Twitter-XLM-RoBERTa. All models were fine-tuned using a customized Focal Loss objective combined with Label Smoothing to mitigate majority class bias and prevent overconfidence. Out-of-fold probability representations from individual models were aggregated through a Logistic Regression stacking meta-classifier (stacking_c1). On the official test set, our system achieved a Macro F1-score of 0.8141 and an Accuracy of 0.8343, demonstrating the effectiveness of combining specialized dialectal representations with loss regularization and meta-learning.

Dialect sentiment analysis asks for the polarity of a sentence written in a regional Arabic variety rather than Modern Standard Arabic. We present a system for Subtask 1 of DialectSentEval built on two observations about the data. First, the task is one of zero-shot transfer rather than classification: the test set introduces Lebanese, a dialect with no labeled training data and the highest out-of-vocabulary rate in the task (52.0% of word types, against 19.5–41.0% for the seen dialects). Second, k-fold cross-validation cannot see this, since it measures generalization to unseen sentences and never to an unseen dialect. We therefore validate with a leave-one-dialect-out diagnostic and score every component on the official leaderboard. Three dialectally diverse encoders (MARBERTv2, CAMeLBERT-DA, AraBERTv2) are trained with Fast Gradient Method adversarial training, then refined by confidence-filtered self-training on the unlabeled test set and character-level noise augmentation. Adversarial training contributes +0.040 Macro F1 over a gold-only ensemble, self-training a further +0.027, and per-class pseudo-label filtering +0.018. The final system reaches 0.8667 Macro F1, ranking 1st of 12 submissions, 4.6 points above the runner-up.

Sentiment swap is the task of rewriting a sentence so that it carries the opposite polarity while its topic stays the same. We present a system for Arabic sentiment swap built on two observations about the data. First, the swap is a local edit rather than a paraphrase: 91% of reference rewrites stay within two words of their source. Second, the polarity to swap away from has to be predicted, because the test data does not provide it. We therefore fine-tune mT5 and refine it with Direct Preference Optimization under a reward that balances target-polarity probability against content preservation. At inference we pool candidates from four generators, one of them a minimal-edit tagger, together with a retrieval index, rerank them with the same balance, and repair them with polarity-flip lexicons mined from the training set. Source polarity is supplied by a dedicated classifier that reaches 97.3% accuracy, against 74.3% for the off-the-shelf sentiment model that the evaluation itself uses. Every component was scored on the official leaderboard: generation-side work contributes +0.044 Sentiment Style Accuracy, and the dedicated polarity classifier a further +0.155. The final system reaches 0.9581 Sentiment Style Accuracy, 56.91 BLEU and 72.18 chrF, ranking 1st of 15 submissions.

We present ArA-DF 2026, the first shared task for Arabic speech deepfake detection, held at ArabicNLP-2026. Unlike existing benchmarks that conflate all distribution shift into a single evaluation set, ArA-DF 2026 separates dialect generalization (Track 1: unseen Arabic dialects and speakers) from acoustic robustness (Track 2: codec compression, noise, reverberation, re-encoding). We release 43,500 labeled training utterances spanning MSA and regional dialects, plus blind evaluation sets of 144,210 (Track 1) and 127,746 (Track 2) utterances, ranking systems by EER. Fourteen teams competed on Track 1 and twelve on Track 2, with eight system description papers submitted. The best submissions cut EER from 14.54% to 0.87% (Track 1) and from 27.11% to 0.66% (Track 2). Across independent submissions, three findings recur: augmentation and data curation move performance far more than architecture; simpler pooling heads often outperform graph-attention back-ends once augmentation is applied; and external synthetic Arabic speech with mismatched generation artifacts hurts rather than helps. Residual errors concentrate on short utterances and on spoofed audio misclassified as bona fide.

This paper presents the SparkShield system submitted to ArA-DF 2026 Track 2, which focuses on robust Arabic speech deepfake detection under realistic acoustic degradations. To improve robustness, we apply a comprehensive augmentation pipeline involving background noise, reverberation, codec compression, reencoding, and volume perturbation. Our system consists of three parallel subsystems using XLS-R models with 300M, 1B, and 2B parameters as front-end feature extractors. Each subsystem employs learnable layer-wise feature fusion, followed by a Multi-kernel gated Convolution (MultiConv) module and Multi-Head Attention Pooling (MHAP) for utterance-level spoof detection. The output logits of the three independently trained subsystems are finally fused. The proposed system achieves an EER of 0.6575% on the evaluation set and ranks first in ArA-DF 2026 Track 2.

This paper describes the system submitted by team PogranichnyZashchitnik to the ArA-DF 2026 shared task on Arabic Audio Deepfake Detection. We build on WavLM Large as a self-supervised feature extractor, fine-tuned with Low-Rank Adaptation (LoRA) so that only a small fraction of parameters need updating. Representations from all 25 hidden layers are blended via a learnable weighted sum, pooled into utterance-level embeddings with Attentive Statistics Pooling, and classified by an OC-Softmax head that treats bonafide speech as the target class. Combined with an augmentation pipeline of codec compression, additive noise, and reverberation, this single system achieves 12.54% EER on Track 1 (Dialect Generalization, 14th/15) and 5.66% EER on Track 2 (Acoustic Robustness, 8th/14), outperforming the organizer baseline on both tracks. Post-evaluation analysis shows that utterances shorter than two seconds are disproportionately difficult, and that a non-trivial fraction of spoofed samples are scored as confidently bonafide under dialect mismatch.

We describe our submission to ArA-DF 2026, the ArabicNLP shared task on Arabic speech deepfake detection, in which systems classify audio as real or deepfake. We participated in both tracks: Track 1 (Dialect Generalization) and Track 2 (Acoustic Robustness). Our systems investigated finetuning three large self-supervised speech encoders — WavLM, wav2vec2-XLS-R, and HuBERT — with a learnable attentive-pooling head. We evaluated each encoder separately and further combined encoders through a logistic-regression meta-classifier. In addition to improve robustness we apply an acoustic degradation pipeline (additive noise, lossy codec re-encoding, band-limiting, and gain) to 50% of the training data. For track one the best achieved result is, a fully fine-tuned WavLM with attentive pooling and augmentation, which scored 2.19% EER test set, and a deepfake-pretrained XLS-R achieves 7.80% EER on Track2 which out performed the official baseline. Our ablation study shows that augmentation and backbone choice are the primary factors, while adding correlated ensemble members yields no gain.

We describe our system for Track 1 (dialectgeneralization) of the ArA-DF-2026 Arabic speech deepfake detection shared task. Starting from the organizers’ XLS-R+AASIST baseline (14.65% EER on the development test set), we show that (i) raw-waveform augmentation (RawBoost) roughly halves the error rate, (ii) replacing the AASIST back-end with a simple mean-pooling head over a posttrained self-supervised frontend removes a hard performance ceiling that persists even after full convergence, and (iii) a large, architecturally diverse ensemble is the only lever that keeps improving once single-model performance saturates. Our final system, a 13-member weighted ensemble spanning three frontend sizes, two back-ends, and four training-data recipes, reaches 2.34% EER on the development-test leaderboard and 2.42% EER on the official held-out blind evaluation set (144,210 utterances, 9× larger than the development test). We also report two negative results of practical value: injecting a large (54k-clip) multi-system Arabic text-to-speech corpus and a vocoder-copy-synthesis pseudospoof corpus both degrade accuracy, indicating that indiscriminately adding synthetic speech is harmful when its generation artifacts do not match the target domain. Our code is publicly available at https://github.com/Body123/aradf-2026-saar-df.

ArA-DF 2026 is a shared task on binary Arabic speech deepfake detection, with two tracks: dialect generalisation and acoustic robustness. For both we submit rank-normalised score fusions of complementary detectors. Track 1 fuses five components in two stages: three fine-tuned XLS-R-300M networks and two ResNet-18 classifiers over linear-prediction residual and group-delay representations. Track 2 fuses two XLS-R-300M detectors differing in augmentation strategy. We reach 2.8961 EER on Track 1 and 2.7626 on Track 2, against baselines of 14.54 and 27.11. Our analysis on the released test labels shows individual EER to be a poor guide to fusion contribution, which depends on error independence: both fusions improve on their best member by over 40%, and components far weaker in isolation than any we submitted improve them further.

We describe Raid AI’s entry to Track 1 (Dialect Generalization) of the ArA-DF 2026 shared task on Arabic speech deepfake detection, which placed 1st of 15 teams at 0.8719% Equal Error Rate against a 14.54% official baseline. The configuration of the submitted system is proprietary and is not disclosed. We instead report the ablation behind it, re-scored on the official 144,210-clip evaluation partition against the labels released after the challenge, with bootstrap confidence intervals and paired tests throughout. On development-test, corpus and augmentation alone move a fixed architecture from 13.13% to 1.37% EER. Encoder scale is not monotonic: a 1B encoder beats a 2B under a recipe matched on corpus, augmentation, learning rate, class weights and effective batch size, though the gap is inside the run-to-run spread we measure from three seeds. Three-crop test-time augmentation improves all ten models tested, by 17.4% relative on average, and is the only effect immune to seed variance because it is measured within a checkpoint; a per-crop decomposition shows the gain comes from averaging rather than from choosing a better window.

Detecting synthetic and voice-converted speech remains difficult for low-resource languages with dialectal diversity, where systems must generalize across regional dialects and unseen acoustic channels. We present the UZH-CL submission to the ArA-DF 2026 Shared Task on Arabic speech deepfake detection, covering Track 1 (dialect generalization) and Track 2 (acoustic robustness). We freeze a W2V-BERT-2.0 backbone and adapt it with Wavelet Prompt Tuning, updating under 1% of parameters, and aggregate multi-layer representations with cross-layer attention and a general attentive-statistics pooling head rather than a specialized graph backend. Complementary detectors are obtained by varying adaptation strategy, training data, augmentation, and encoder family. We find that the two shift types require different fusion regimes: a broad multi-window ensemble for dialect generalization, and a compact, channel-matched, center-crop ensemble for acoustic robustness. Official evaluation yields 1.96% EER on Track 1 (6th place) and 1.04% EER on Track 2 (3rd place), corresponding to 87% and 96% relative reductions over the XLS-R+AASIST baselines.

This paper presents our submission to the ArADF 2026 Shared Task on Arabic Audio Deepfake Detection, focusing on Track 1, which evaluates robustness to previously unseen Arabic dialects. Our proposed system combines two complementary self-supervised speech encoders, XLS-R-300M and WavLM-Large, each integrated with an AASIST graph-attention back-end for spoofing detection. The final prediction is obtained through score-level fusion of the two independently trained models, enabling the system to leverage the complementary representations learned by each encoder. To improve robustness to domain shift, we introduce two complementary adaptation strategies: telephony channel augmentation to reduce the mismatch between the training and evaluation domains, and an iterative pseudolabel self-training framework that exploits highconfidence predictions on unlabeled evaluation data for progressive model refinement. On the official ArA-DF 2026 Track 1 evaluation set, our system achieved an Equal Error Rate (EER) of 0.9265%, ranking 4th on the official leaderboard. These results demonstrate that combining complementary self-supervised speech representations with domain-aware adaptation techniques provides an effective and robust solution for cross-dialect Arabic audio deepfake detection.

Function calling connects language models to external tools, but Arabic structured tool use remains insufficiently evaluated. We present , a shared task at ArabicNLP 2026 built around 12,220 Arabic requests spanning five dialect groups and eight service domains. Each request has four candidate schemas drawn from a registry of 27 tools, including seven tools that never occur as gold calls; 223 requests require abstention. Systems decide whether to call, select a function, and extract its arguments. Two ranked tracks evaluate core calling and reasoning-trace compliance, with a third diagnostic examining dialect robustness. Twenty-two teams submitted to the blind phase, and 21 produced valid predictions on 1,125 test requests. We analyze 73 submissions and 13 system descriptions. A LoRA-adapted 4B model ranks first on both competition tracks, with scores of 0.9279 and 0.9400, while argument exact match peaks at 0.8816. However, a deterministic rule combining candidate order and released metadata matches the best observed function-name accuracy of 0.9973, limiting what this metric establishes about language understanding. Exact argument extraction provides most of the competitive separation. Maghrebi has the lowest score for every system, although its small test subset and differences in example composition limit interpretation. We provide public task data, a baseline checkpoint, the scorer, and final rankings, and document evaluation shortcuts and remaining reproducibility limitations.[<https://huggingface.co/spaces/TuwaiqAcademy/AISA-ArabicFC-Shared-Task>]

We describe the Saudi Legal FC submission to the AISA-ArabicFC shared task on Arabic function calling. Our system is a parameter-efficient fine-tune of Qwen2.5-7B-Instruct with LoRA, trained with completion-only loss masking. Beyond the model, our central contribution is methodological: we use written, falsifiable score predictions and a validated local scorer to gate submissions, with a post-processing safety net that uses no gold at inference. On the public development set our best configuration reaches Argument Exact-Match (ArgEM) 0.828 and function-name accuracy 1.0 (Overall-B 0.914) under the official v1.3 evaluator. Its weights were then lost; the blind submission comes from a reconstruction selected against that model’s frozen predictions, and ranks 12th on Track B with Overall-B 0.8946. We also report two negative findings—one retraining run, one rejected hypothesis—and an artifact-identity failure our behavioral checks caught before submission.

We describe the Durham-Jazan submission to AISA-ArabicFC 2026, an Arabic function-calling task in which a system selects one tool from four candidates and fills its arguments, scored on tool choice (FnAcc) and on strict exact match of the whole argument dictionary (ArgEM). Across six QLoRA fine-tunes of ALLaM-7B and Qwen2.5-7B, trained on the provided data, negatives derived from it and 68 hand-written no-tool utterances, and run on one 16 GB GPU, we vote on each argument separately, then apply two deterministic rules estimated from training statistics. The system reaches ArgEM 0.824 on development under the final evaluator and 0.826 on the blind test, placing 9th of 17 in Track A, 8th of 17 in Track B and 11th of 21 in Track C. Our main findings concern the benchmark. The reference tool is listed first on every training and development row that calls one, in every data release. Shuffling the candidates at inference costs the voted system 0.202 FnAcc and 0.190 ArgEM, and retraining one adapter per base on permuted candidates removes the dependence with no loss of ArgEM. One post-processing rule contributes 0.028 ArgEM and is the only component whose gain separates from zero, the other targets a convention the final data release removed and is inert, and the largest effect we can measure comes from the benchmark’s construction rather than from our modelling.

We describe NABIQ, the system submitted by team Wajeh (واجه) to Tracks A and B of the AISA-ArabicFC shared task on Arabic function calling. NABIQ contains no large language model, no neural network of any kind, no external API and no GPU. It is a CPU-based pipeline of classical ML, retrieval and rule-based repair: a TF-IDF router, a per-tool nearest-argument retriever over train, span extractors with train-mined gazetteers, and a no-call detector. Its core mechanism is an evidence gate: a candidate value may replace a prediction only if it is a normalised substring of the utterance, a member of a closed canonical map, or a date whose tokens occur in the text. On the blind test the frozen system scored Overall A 0.7904 (rank 16) and Overall B 0.8282 (rank 15) at ThinkRate 1.0000. We keep five evaluation levels distinct because the submitted system drops 0.21 ArgEM from development to test. A rehearsal on 545 held-out training rows, run before the blind release, tracked our official Overall A far better than the tuned development estimate (composite error 0.0033 against 0.1724, cancelling larger FnAcc and ArgEM errors); that contrast, not the components, is our result.

Tool calling has become increasingly important for agentic systems, but remains challenging in multilingual settings, particularly for Arabic. We present our system for the AISA-ArabicFC shared task at ArabicNLP 2026. Our approach combines decomposed multi-task fine-tuning, function-centric few-shot augmentation, and hybrid schema-aware inference. On the development set, our system achieves a function-name accuracy of 1.000 and a normalized argument exact match of 0.858. On the blind test set, the released baseline re-scored with the final scorer obtains a Track A score of 0.8000, compared with 0.9111 for our submission. Our system achieves a function-name accuracy of 0.9973 and an argument exact match score of 0.8536 on this set, ranking seventh in Track A. Our solution is released in the GitHub repository[<https://github.com/phud774/PuDyDG-AISA-ArabicNLP>].

We describe the QMUL_AK submission to AISA-ArabicFC 2026. The system adapts the Arabic-English ALLaM-7B-Instruct model to function calling with low-rank adaptation (LoRA). It combines an Arabic schema-aware prompt, dialect and no-call resampling, a token-weighted training objective, and deterministic validation of generated calls against the candidate tool schema. The same output supports Track A function calling and Track B reasoning-augmented calling. Training uses only the official task data. On the development set, the system achieves 0.9982 function accuracy and 0.8920 argument exact match, corresponding to overall A and B scores of 0.9345 and 0.9454. On the blind test set, the system reaches overall A and B scores of 0.9253 and 0.9380, ranking 3rd on both submission tracks and 1st of 21 teams on the Track C dialect-robustness diagnostic with a macro-averaged score of 0.9075 and the smallest dialect gap in the shared task (0.077). The results show that function selection is nearly saturated, while argument extraction remains the main source of error.

We describe Abjad’s submission to the AISA-ArabicFC shared task (Track B), which targets Arabic function calling with explicit natural-language reasoning prior to the structured call. Rather than fine-tuning, our system builds a robust inference and evaluation pipeline around the organizers’ baseline model, AISA-AR-FunctionCall-Think, diagnosing and fixing two pipeline-level bugs that were silently depressing scores. On the 545-example development set, our system reaches a function-name accuracy of 96.15% but an argument exact-match of only 40.6%, yielding an official Track B score (overall_b) of 0.6665, below the organizer-reported baseline of 0.739. We present a fine-grained error analysis showing that failures concentrate almost entirely in argument value formatting, including currency codes, ISO language codes, and transliteration, rather than in tool selection or task comprehension. These findings suggest that targeted fine-tuning or post-processing for structured-value normalization is a promising direction for improving Arabic function-calling performance. abjad aisa arabicfc.pdf

We describe Siraj’s Track B submission to AISA-ArabicFC. A QLoRA-tuned ALLaM-7B is followed by a separately trained argument repairer whose edits pass a type-restricted acceptance filter, then by one deterministic year rule. On the blind test set (1,125 rows) it scores 0.8245 ArgEM and 0.9101 Track B, 10th of 17 systems; the post-editing stage is worth about one example on test against four on development. Every rule that error inspection suggested was adjudicated on out-of-fold training predictions, scored with the task’s own published matcher, before it could touch development predictions; all five were rejected, one of them breaking 168 more rows than it fixed. Two development-time conclusions did not survive: the top three systems finished 5.5 to 5.7 ArgEM points ahead, so our error rate was not a task-level ceiling, and a later dataset revision lowers our development ArgEM from 0.850 to 0.8080 and neutralises our strongest rule. Post-editing rules fitted to gold conventions can inherit changes in those conventions.

We present ArgTune, a modular pipeline for Arabic function calling developed for the AISA-ArabicFC shared task. ArgTune separates tool selection from argument extraction using three LoRA-fine-tuned AISA-AR-FunctionCall-Think (270M) models: a tool-selection model, a general argument model, and a specialist model for seven challenging tools, allowing each model to focus on a more specific task. Targeted augmentation, data balancing, and query-grounded post-processing further address difficult argument cases. On the official test set, ArgTune achieved 0.9964 FnAcc, 0.8506 ArgEM, and an overall score of 0.9089, ranking 8th. The results highlight the remaining difficulty of exact argument extraction despite highly accurate tool selection.

Arabic function calling requires models to identify appropriate tools and reproduce arguments in the exact representation expected by evaluators. This is challenging due to dialectal variation, lexical alternatives, and surface-form normalization. We present a hybrid system based on Qwen2.5-7B-Instruct adapted with 4-bit QLoRA, followed by conservative deterministic post-processing. The neural model generates flexible tool calls, while a lightweight deterministic layer parses, validates, and casts the output to improve compliance with the exact-match evaluation format. On the official blind evaluation, our system obtains 93.24% Function Accuracy and 58.38% Argument Exact Match (Track A: 72.32%, Track B: 71.37%). Error analysis indicates that many remaining failures arise from argument-surface mismatches rather than function-selection errors, supporting a design where generative modeling handles semantics while deterministic processing controls exact realization.

We describe the Team Yalla submission to the AISA-ArabicFC shared task on Arabic function calling, which ranked second in both tracks (0.9273 Track A, 0.9395 Track B), within 0.0006 of the winning system. The headline metric, exact argument match (ArgEM), is unforgiving: a single wrong surface form loses the whole example. The provided targets bury the scored object inside a rationale: the tool name and argument dictionary are only 30.3% of the target characters. We trained on the 30.3%, which is our largest single gain (+0.020 ArgEM in a matched comparison). A matched follow-up shows that reasoning need not be removed entirely: emitting it after the arguments trails the JSON-only target by only 0.008 ArgEM. We add a reinforcement-learning stage that uses the official scorer as the reward, and combine three deliberately diverse fine-tunes with a per-argument majority vote, which repairs individual argument slots rather than choosing whole predictions. Every member is a LoRA fine-tune of a public 4B or 7B Qwen model trained on the official split alone, with no distillation and no external data. We note that two of the three voting members emit no rationale, so our Track B think field is attached after prediction from a separate reasoning-trained model whose predicted call for that row is identical: it is a post-hoc explanation, not the computation that produced the arguments. We report the experiments that failed alongside those that worked.

We describe Ṣaqr, ranked first on both scored tracks of the AISA-ArabicFC shared task on Arabic function calling, scoring 0.9279 on Track A and 0.9400 on Track B over a 1,125- instance blind test set. The system is deliber- ately plain: one Qwen3-4B model with LoRA, two epochs, provided data only. Our contri- bution is not the system but an accounting of where score comes from. We find that scale helped only up to a point in our single-run, cross-family comparison, and that the remain- ing gains came instead from inference-time de- cisions worth +0.0063 Overall A and +0.0223 Overall B (Appendix A) with no retraining. prefilling the reasoning trace so the model can- not skip it, and parsing arguments against the declared schema . We also report two proper- ties of the data that bound any system’s score: abstention is decidable from prompt structure without reading the query, and the gold anno- tation is self-inconsistent on precisely the slots that separate strong systems.

Argument exact-match (ArgEM) dominates the AISA-ArabicFC score. Under one prompt, a base gemma-4-12B-it reaches 0.9853 FnAcc but only 0.6940 ArgEM on dev, suggesting that argument convention rather than tool selection is the binding constraint. Fine-tuning on the provided train split raises ArgEM to 0.8880 while FnAcc increases only from 0.9853 to 0.9945, yet the model still fails 8.4% of the 10,500 positive training rows. Across rules, retrieval, prompting, selection, reranking, and fusion, dev-selected gains did not consistently transfer beyond development; smaller gains from other tested interventions remain unresolved. Explicit convention cards add 7.5 ArgEM points on a dev-selected 228-row subset for an unconfigured model, but give no measurable gain to the already configured pipeline. A scorer-equivalent gold answer occurs in a 20-sample pool for 34/56 (60.7%) rows missed by the raw fine-tune, but our selectors cannot exploit it. Our official prompt-only gemma-4-31B-it submission placed 6th of 17 on Track A, 5th of 17 on Track B, and 4th of 21 on Track C. The 12B fine-tune used for analysis was developed first but inadvertently omitted from the blind-test submission, so it has no official test score.

Reliable function calling is challenging for Arabic due to dialectal variation and limited task-specific resources. We present Makeen, a data-centric approach for the AISA-ArabicFC shared-task, and extend it with span extraction and confidence-gated routing. The best post-submission variant achieves 83.20% argument exact match and 91.11% Track B on development data under held-out (cross-validated) routing, while the submitted model achieves 82.65% argument exact match and 91.06% Track B on the blind test set, ranking 9th in Task B. We also release two Arabic tool-use datasets and show that external-tool augmentation improves unseen-tool argument exact match from 44.49% to 66.53%. These results highlight the importance of data quality, argument grounding, and tool-level generalization for Arabic function calling.

Arabic Function Forge is our submission to AISA-ArabicFC 2026. Our system is training-free: it uses a single off-the-shelf open-weights model, prompted once per request, with no parameter updates. It placed fifth on Track A, fourth on Track B, and sixth on Track C. Our central observation is that function selection is nearly saturated on this benchmark while argument extraction is not: we reach 0.9893 function accuracy but only 0.8686 argument exact match, so 94.9% of the remaining headroom lies in arguments. The dominant failure is not misunderstanding but over-interpretation: the model translates values, converts dates to ISO form, and adds fields the user never provided. Our system is designed to suppress this behaviour by preserving argument values as written. We further find that the dialect penalty is concentrated rather than graded, with most of the drop occurring on Maghrebi.