Entity-level unlearning is usually evaluated by what a model says: whether it stops naming the target, refuses a query, or shifts a Truth Ratio distribution. These output-level tests, however, do not show whether a subject’s internal representation has been attenuated. We introduce the Entity Representation Unlearning Framework (ERUF), a representation-aware framework that mines subject-specific activation signatures, suppresses the corresponding activation direction, and distills the behavior into LoRA parameters. Among evaluated baselines, ERUF is the only method that jointly achieves surface level suppression, internal attenuation, and utility preservation. On TOFU forget10, ERUF achieves FQ =0.99 and MU =0.62, matching reported oracle utility while approaching oracle forget quality. Across most standard foundation-model settings, ERUF maintains low leakage and low internal target activation (SMR =0.00–1.10%, EL10 <0.06, utility drift <3%). On Llama-3.1-8B, adversarial entity recovery falls from 63.89% to 20.15%, while name-agnostic recovery decreases by 72.7–77.4%. Joint surface/internal diagnostics further reveal scale-dependent behavior in reasoning-prior models that surface metrics alone would miss. We interpret these results as operational evidence of representation-level attenuation, not as a formal guarantee of irreversible deletion. Code is available in the repository.

Function routing—selecting the correct API call from a fixed catalog given a natural-language request — is a deployment problem where small students are attractive but knowledge distillation gains are typically reported single-seed, at scales where seed variance is unknown. On a 740-instance healthcare API routing task with a 1.5B Qwen student and a 20B teacher, we compare eight KD variants against supervised cross-entropy, using three to six seeds for key configurations. We find: (i) per-seed standard deviation ranges from 2.8 to 48.7 percentage points, swallowing every claimed KD gain below five points; (ii) three of seven KD variants exhibit bimodal collapse, with at least one in three to five seeds falling below 55% accuracy while the others train normally, and a fourth showing elevated variance; (iii) collapse has distinct modes — wrong-function selection for ce_Knowledge Distillation and ce_paraphrase, and a previously undocumented output-truncation mode for reasoning_kd, where the model emits reasoning but terminates before producing a function name (0.9% accuracy); (iv) only progressive_kd and rank_kd avoid collapse across observed seeds, with σ ≤ 3.9 pp; (v) a naive cross-split +3.78 pp gain from input enrichment reverses to −2.70 pp under controlled within-split multi-seed re-testing. Single-seed evaluation is therefore unable to detect central failure modes in small-model KD.

Large language models (LLMs) frequently exhibit performance biases against regional dialects of low-resource languages. However, frameworks to quantify these disparities remain scarce. We propose a two-phase framework to evaluate dialectal bias, operationalized as comprehension degradation relative to standard Bengali, in LLM question-answering across nine Bengali dialects. First, we translate and gold-label standard Bengali questions into dialectal variants adopting a retrieval-augmented generation (RAG) pipeline to prepare 4,000 question sets. Since traditional translation quality evaluation metrics fail on unstandardized dialects, we evaluate fidelity using an LLM-as-a-judge, which human correlation confirms outperforms legacy metrics. Second, we benchmark 19 LLMs across these gold-labeled sets, running 68,395 RLAIF evaluations validated through multi-judge agreement and human fallback. Our findings reveal severe performance drops linked to linguistic divergence. For instance, responses to the highly divergent Chittagong dialect score 5.44/10, compared to 7.68/10 for Tangail. Furthermore, increased model scale does not consistently mitigate this bias. We contribute a validated translation quality evaluation method, a rigorous benchmark dataset, and a Critical Bias Sensitivity (CBS) metric for safety-critical applications.

Gender bias in LLMs has been studied extensively in model outputs, with biased prompts shown to amplify stereotyped generations. Whether such bias propagates into text produced by humans who use these systems, however, remains underexplored. We investigate whether gender bias in an LLM writing assistant transfers into career plan essays written by students. We first verify that a gender-biased prompt induces gender-differentiated language in LLM-generated essays, while a neutral prompt does not. We then recruited participants (N = 123) in a controlled environment to write career plan essays for paired biographical profiles differing only in gender under three conditions: no AI assistance, neutral LLM assistance, or gender-biased LLM assistance. Students in the biased condition produced essays with a significantly larger agentic gap and more gender-stereotypic occupation suggestions than those in the control and neutral conditions. Our results also reveal that this bias transfer is asymmetric: agency is suppressed in female-target essays while male-target writing remains largely unaffected. Our findings highlight the risk of bias propagation in AI-assisted writing, calling for fairness-aware design in educational AI tools.

Large language models (LLMs) show promise in generating supportive responses for mental health and counseling applications. However, their responses often lack cultural sensitivity, contextual grounding, and clinically appropriate guidance. This work addresses the gap of how to systematically incorporate domain-specific, clinically validated knowledge into LLMs to improve counseling quality. We utilize and compare two approaches, retrieval-augmented generation (RAG) and a knowledge graph (KG)–based method, designed to support para-counselors. Our KG is constructed manually and clinically validated, capturing causal relationships between stressors, interventions, and outcomes, with contributions from multidisciplinary people. We evaluated multiple LLMs in both settings using BERTScore F1 and SBERT cosine similarity, as well as human evaluation across five metrics, which is designed to directly measure the effectiveness of counseling beyond similarity at the surface level. The results show that KG-based approaches consistently improve contextual relevance, clinical appropriateness, and practical usability compared to RAG alone, demonstrating that structured, expert-validated knowledge plays a critical role in addressing LLMs limitations in counseling tasks.

Despite the growing progress in Natural Language Inference (NLI) research, resources for the Bengali language remain extremely limited. Existing Bengali NLI datasets exhibit several inconsistencies, including annotation errors, ambiguous sentence pairs, and inadequate linguistic diversity, which hinder effective model training and evaluation. To address these limitations, we introduce BNLI, a refined and linguistically curated Bengali NLI dataset designed to support robust language understanding and inference modeling. The dataset was constructed through a rigorous annotation pipeline emphasizing semantic clarity and balance across entailment, contradiction, and neutrality classes. We benchmarked BNLI using a suite of state-of-the-art transformer-based architectures, including multilingual and Bengali-specific models, to assess their ability to capture complex semantic relations in Bengali text. The experimental findings highlight the improved reliability and interpretability achieved with BNLI, establishing it as a strong foundation for advancing research in Bengali and other low-resource language inference tasks. The link to the BNLI dataset: https://github.com/FarahHaque/BNLI-Dataset.git

Attention sinks assign large, largely content-independent attention to the first position. Ran-Milo et al. (2026) explain GPT-2’s sink with a parameter-level circuit: a learned query bias 𝐛_Q aligns with the key projection of the first position’s effective positional encoding, whose few massive coordinates amplify the score. They infer that other architectures build sinks differently but do not measure this claim. We reproduce the original intervention effects and stress-test their causal interpretation. Baseline plus nine interventions and four structural analyses reproduce under the original evaluation, with two scope qualifications; the massive coordinates are exactly 138, 378, 447. Yet nullifying 𝐛_Q leaves 44.5% of the sink, which follows an activation-dependent query route into the same first-position key. This *anchor*–*route* distinction unifies the extensions: with GPT-2 scale, the bias route fades while the activation-dependent route concentrates into fewer carrier heads; across 16 checkpoints in four families, no qualitative profile over mutually available conditions spans all families; and across five pre-training runs, anchoring recurs despite unstable coordinate identities. In GPT-2-small at 40 tokens, Swap EPE incurs about one quarter of the cross-entropy cost of Remove First PE. The stable phenomenon is an anchoring computation, not a universal parameter-level circuit, so mitigation must be validated against the target model’s wiring.

Conventional research on speech recognition modeling relies on the canonical form for most low-resource languages while automatic speech recognition (ASR) for regional dialects is treated as a fine-tuning task. To investigate the effects of dialectal variations on ASR we develop a 78-hour annotated Bengali Speech-to-Text (STT) corpus named Ben-10. Investigation from linguistic and data-driven perspectives shows that speech foundation models struggle heavily in regional dialect ASR, both in zero-shot and fine-tuned settings. We observe that all deep learning methods struggle to model speech data under dialectal variations, but dialect specific model training alleviates the issue. Our dataset also serves as a out-of-distribution (OOD) resource for ASR modeling under constrained resources in ASR algorithms. The dataset and code developed for this project are publicly available.

Recent progress in NLP research has demonstrated remarkable capabilities of large language models (LLMs) across a wide range of tasks. While recent multilingual benchmarks have advanced cultural evaluation for LLMs, critical gaps remain in capturing the nuances of low-resource cultures. Our work addresses these limitations through a Bengali Language Cultural Knowledge (BLanCK) dataset including folk traditions, culinary arts, and regional dialects. Our investigation of several multilingual language models shows that while these models perform well in non-cultural categories, they struggle significantly with cultural knowledge and performance improves substantially across all models when context is provided, emphasizing context-aware architectures and culturally curated training data.

This study proposes a hybrid framework for Bangla-to-Python code generation, emphasizing improved code accuracy through a two-phase pipeline: generation and debugging. During development, standalone models such as TigerLLM and StarCoder achieved modest accuracies of 27% and 24%, respectively, while more advanced models, Gemini-1.5-flash and Gemma, reached 60% and 64%. Integrating Gemma with the gpt-oss debugger substantially increased accuracy to 99.75%, highlighting the critical role of a dedicated debugging stage. In testing on unseen data, gpt-oss alone achieved 67%, which improved to 71% with self-debugging. The highest performance, 84%, was obtained by pairing Gemini-2.5-flash as the generator with gpt-oss for debugging. These findings demonstrate that combining a strong generative model with an effective debugging component yields superior and robust code generation results, outperforming existing approaches such as TigerLLM. The full implementation of the framework is publicly available at https://github.com/fahmid-juboraj/Code_generation.

Translating from a standard language to its regional dialects is a significant NLP challenge due to scarce data and linguistic variation, a problem prominent in the Bengali language. This paper proposes and compares two novel RAG pipelines for standard-to-dialectal Bengali translation. The first, a Transcript-Based Pipeline, uses large dialect sentence contexts from audio transcripts. The second, a more effective Standardized Sentence-Pairs Pipeline, utilizes structured local_dialect:standard_bengali sentence pairs. We evaluated both pipelines across six Bengali dialects and multiple LLMs using BLEU, ChrF, WER, and BERTScore. Our findings show that the sentence-pair pipeline consistently outperforms the transcript-based one, reducing Word Error Rate (WER) from 76% to 55% for the Chittagong dialect. Critically, this RAG approach enables smaller models (e.g., Llama-3.1-8B) to outperform much larger models (e.g., GPT-OSS-120B), demonstrating that a well-designed retrieval strategy can be more crucial than model size. This work contributes an effective, fine-tuning-free solution for low-resource dialect translation, offering a practical blueprint for preserving linguistic diversity.

Writing systems of Indic languages have orthographic syllables, also known as complex graphemes, as unique horizontal units. A prominent feature of these languages is these complex grapheme units that comprise consonants/consonant conjuncts, vowel diacritics, and consonant diacritics, which, together make a unique Language. Unicode-based writing schemes of these languages often disregard this feature of these languages and encode words as linear sequences of Unicode characters using an intricate scheme of connector characters and font interpreters. Due to this way of using a few dozen Unicode glyphs to write thousands of different unique glyphs (complex graphemes), there are serious ambiguities that lead to malformed words. In this paper, we are proposing two libraries: i) a normalizer for normalizing inconsistencies caused by a Unicode-based encoding scheme for Indic languages and ii) a grapheme parser for Abugida text. It deconstructs words into visually distinct orthographic syllables or complex graphemes and their constituents. Our proposed normalizer is a more efficient and effective tool than the previously used IndicNLP normalizer. Moreover, our parser and normalizer are also suitable tools for general Abugida text processing as they performed well in our robust word-based and NLP experiments. We report the pipeline for the scripts of 7 languages in this work and develop the framework for the integration of more scripts.

People are expressing their opinions online for a lot of years now. Although these opinions and comments provide people an opportunity of expressing their views, there is a lot of hate speech that can be found online. More specifically, sexist comments are very popular affecting and creating a negative impact on a lot of women and girls online. This paper describes the approaches of the SemEval-2023 Task 10 competition for Explainable Online Sexism Detection (EDOS). The task has been divided into 3 subtasks, introducing different classes of sexist comments. We have approached these tasks using the bert-cased and uncased models which are trained on the annotated dataset that has been provided in the competition. Task A provided the best F1 score of 80% on the test set, and tasks B and C provided 58% and 40% respectively.

Topic modeling of a text corpus is one of the most well-studied areas of information retrieval and knowledge discovery. Despite several decades of research in the area that begets an array of modeling tools, some common problems still obstruct automated topic modeling from matching users’ expectations. In particular, existing topic modeling solutions suffer when the distribution of words among the underlying topics is uneven or the topics are overlapped. Furthermore, many solutions ask the user to provide a topic count estimate as input, which limits their usefulness in modeling a corpus where such information is unavailable. We propose a new topic modeling approach that overcomes these shortcomings by formulating the topic modeling problem as a community detection problem in a word association graph/network that we generate from the text corpus. Experimental evaluation using multiple data sets of three different types of text corpora shows that our approach is superior to prominent topic modeling alternatives in most cases. This paper describes our approach and discusses the experimental findings.

Intentionally luring readers to click on a particular content by exploiting their curiosity defines a title as clickbait. Although several studies focused on detecting clickbait titles in English articles, low-resource language like Bangla has not been given adequate attention. To tackle clickbait titles in Bangla, we have constructed the first Bangla clickbait detection dataset containing 15,056 labeled news articles and 65,406 unlabelled news articles extracted from clickbait-dense news sites. Each article has been labeled by three expert linguists and includes an article’s title, body, and other metadata. By incorporating labeled and unlabelled data, we finetune a pre-trained Bangla transformer model in an adversarial fashion using Semi-Supervised Generative Adversarial Networks (SS-GANs). The proposed model acts as a good baseline for this dataset, outperforming traditional neural network models (LSTM, GRU, CNN) and linguistic feature-based models. We expect that this dataset and the detailed analysis and comparison of these clickbait detection models will provide a fundamental basis for future research into detecting clickbait titles in Bengali articles.

Many areas, such as the biological and healthcare domain, artistic works, and organization names, have nested, overlapping, discontinuous entity mentions that may even be syntactically or semantically ambiguous in practice. Traditional sequence tagging algorithms are unable to recognize these complex mentions because they may violate the assumptions upon which sequence tagging schemes are founded. In this paper, we describe our contribution to SemEval 2022 Task 11 on identifying such complex Named Entities. We have leveraged the ensemble of multiple ELECTRA-based models that were exclusively pretrained on the Bangla language with the performance of ELECTRA-based models pretrained on English to achieve competitive performance on the Track-11. Besides providing a system description, we will also present the outcomes of our experiments on architectural decisions, dataset augmentations, and post-competition findings.

Recently BERT has achieved a state-of-the-art performance in temporal relation extraction from clinical Electronic Medical Records text. However, the current approach is inefficient as it requires multiple passes through each input sequence. We extend a recently-proposed one-pass model for relation classification to a one-pass model for relation extraction. We augment this framework by introducing global embeddings to help with long-distance relation inference, and by multi-task learning to increase model performance and generalizability. Our proposed model produces results on par with the state-of-the-art in temporal relation extraction on the THYME corpus and is much “greener” in computational cost.

Incivility in public discourse has been a major concern in recent times as it can affect the quality and tenacity of the discourse negatively. In this paper, we present neural models that can learn to detect name-calling and vulgarity from a newspaper comment section. We show that in contrast to prior work on detecting toxic language, fine-grained incivilities like namecalling cannot be accurately detected by simple models like logistic regression. We apply the models trained on the newspaper comments data to detect uncivil comments in a Russian troll dataset, and find that despite the change of domain, the model makes accurate predictions.

Health support forums have become a rich source of data that can be used to improve health care outcomes. A user profile, including information such as age and gender, can support targeted analysis of forum data. But users might not always disclose their age and gender. It is desirable then to be able to automatically extract this information from users’ content. However, to the best of our knowledge there is no such resource for author profiling of health forum data. Here we present a large corpus, with close to 85,000 users, for profiling and also outline our approach and benchmark results to automatically detect a user’s age and gender from their forum posts. We use a mix of features from a user’s text as well as forum specific features to obtain accuracy well above the baseline, thus showing that both our dataset and our method are useful and valid.