We present H-RAG, our submission to SemEval-2026 Task 8 (MTRAGEval), addressing both Task A (Retrieval) and Task C (Generation with Retrieved Passages). Task A evaluates standalone retrieval quality, while Task C assesses end-to-end retrieval-augmented generation (RAG) in multi-turn conversational settings, requiring both accurate answer generation and faithful grounding in retrieved evidence. Our approach implements a hierarchical parent–child RAG pipeline that separates fine-grained child-level retrieval from parent-level context reconstruction during generation. Documents are segmented into overlapping sentence-based child chunks, while full documents are preserved as parent units to provide coherent context. weighting, and embedding-based similarity rescoring over child chunks. Retrieved evidence is aggregated at the parent level and supplied to an instruction-tuned language model for response generation. H-RAG achieves an nDCG@5 score of 0.4271 on Task A and a harmonic mean score of 0.3241 on Task C (RBagg: 0.2488, RLF: 0.2703, RBllm: 0.6508), underscoring the importance of retrieval configuration and parent-level aggregation in multi-turn RAG performance.

Generating summaries from chart metadata in Arabic presents unique challenges at the intersection of cross-lingual transfer and data-to-text generation. Chart-to-text benchmarks have advanced English-language research, yet Arabic remains without a comparable resource, underscoring its continued underrepresentation in NLP. To cover this gap, we construct the first Arabic ChartSumm benchmark by translating chart metadata and reference summaries from English into Modern Standard Arabic (MSA). Two high-quality machine translation models with contrasting architectures are employed: NLLB-200-distilled-600M, designed for low-resource coverage, and Qwen2.5-1.5B, an open large language model with general multilingual capabilities. A central contribution of this work is a translation quality evaluation that systematically assesses both systems using BLEU, chrF, COMET_ref, and COMET_QE metrics against a Google-Translate Arabic pivot. Results demonstrate that NLLB achieves markedly higher lexical and semantic fidelity. Building on this foundation, we fine-tune two models, mT5 (multilingual) and CAMeL-Lab’s AraBART (Arabic-specific), to generate Arabic summaries from structured chart metadata. Experimental results show that AraBART trained on NLLB translations outperforms other configurations, achieving ROUGE-L = 63.8 and BLEU = 33.1, highlighting the strong dependency of downstream summarization quality on translation accuracy and demonstrating its superior capacity for Arabic generation.

When LLMs are given contradictory passages, the generated response may appear decisive while masking how internal signals behave under conflicting evidence. Does it favor one passage, blend both, or ignore both? We study model behavior under inter-context contradiction by introducing a white-box tracing framework that extracts internal model signals based on hidden-state shifts, counterfactual attribution via passage removal, and token-level uncertainty during generation under two single-passage prompts and one dual-passage prompt with conflicting evidence. We evaluate four open-weight LLMs from the LLaMA (8B and 70B), OLMo, and Qwen on WikiContradict, a human-annotated QA benchmark with contradictory passages. Each dual-passage answer is labeled as committing to passage A or passage B, blending both, or fabricating an answer unsupported by either passage. Comparing these labels with the extracted signals shows that larger models reduce fabrication, but hidden-state and probability-level signals can disagree with the final answer. Moreover, fabricated answers often remain high-confidence, showing that confidence alone is unreliable for identifying ungrounded behavior under contradictory evidence. We further validate on ConflictBank, a larger misinformation-conflict QA dataset, showing consistent disagreement patterns across models. Finally, we test whether pre-generation white-box signals are informative beyond interpretation by training classifiers to predict output behavior. Classifiers using white-box signals outperform probability baselines, showing these signals contain diagnostic information about model behavior under contradictory evidence beyond log probabilities.

Detecting spans of hallucination in LLM-generated answers is crucial for improving factual consistency. This paper presents a span-level hallucination detection framework for the SemEval-2025 Shared Task, focusing on English and Arabic texts. our approach integrates Semantic Role Labeling (SRL) to decompose the answer into atomic roles, which are then compared with a retrieved reference context obtained via question-based LLM prompting. Using a DeBERTa-based textual entailment model, we evaluate each role’s semantic alignment with the retrieved context. The entailment scores are further refined through token-level confidence measures derived from output logits, and the combined scores are used to detect hallucinated spans. Experiments on the Mu-SHROOM dataset demonstrate competitive performance. Additionally, hallucinated spans have been verified through fact-checking by prompting GPT-4 and LLaMA. Our findings contribute to improving hallucination detection in LLM-generated responses.