Large language models increasingly justify legal decisions by naming the statute or precedent behind a verdict, treated as evidence the decision follows from it. We test this directly: holding case facts fixed, we substitute the named legal authority for an unrelated one and decode a model’s evolving verdict from its hidden states. Across seven open-weight models (8B–70B) and four benchmarks spanning judicial and contractual reasoning, when explicitly required to justify a verdict by naming the governing authority, models name the correct one in 66.7%–100% of generations, while the verdict changing when the authority changes is far less consistent: 0.0%–21.7% on CaseHOLD, 30.0%–76.7% on ECHR and SCOTUS, and 43.3%–50.0% on ContractNLI. Neither scale nor a purpose-built legal-reasoning model (a best-effort LoRA reproduction; Section 6) closes this gap. A red-teaming evaluation on five core models finds compliance with an adversarial instruction hidden in the case facts (73.3%–96.4%) exceeds verdict-swap sensitivity by a wide margin, holding without exception regardless of cross-model ranking. Naming a legal authority is thus a poor proxy for a verdict’s dependence on it, and the same verdict remains separately vulnerable to adversarial manipulation; both findings replicate across checks ruling out prompt-wording noise and confounded sampling, and bear directly on using generated legal explanations as compliance or audit artefacts.

Large language model compliance systems are deployed on the assumption that a verdict depends on the regulatory rule it is given. We test this directly across five models and 20 regulatory and platform-policy domains: delete, swap, or negate the governing rule while holding the case fixed, and check whether the verdict changes (OCS) or the model’s internal representation of compliance shifts at all (ICS-delta). Neither moves much: models’ verdicts are often invariant to substantial perturbations of the supplied rule, and the guard model, evaluated here under a custom-rule adaptation of its native taxonomy, is the least rule-sensitive and least accurate of the five, barely above chance (51%, versus 90–92% for general-purpose models). This reflects easy cases more than blanket neglect: on cases where deleting the rule changes a previously correct model prediction, models do track it closely. Neither better prompting nor direct intervention on the model’s internal representations closes this gap. Accuracy alone does not establish that a compliance verdict is grounded in the supplied rule.

Synthetic post-training pipelines commonly filter generated samples with reward models or holistic LLM judges, often relying on two assumptions: that a sample’s faithfulness can be judged without the exact source that produced it, and that rejected samples can be discarded rather than systematically recovered. We present a controlled study of both assumptions across gate configurations, recovery strategies, and generator scales, using adversarially injected corpora to provide ground-truth failure labels. We find that grounding faithfulness judgments in the exact source evidence improves over post-hoc retrieval for stronger judges and substantially outperforms reward-only filtering; hallucination and reward gates reject largely distinct sample populations supporting the need for both; and adaptive recovery, which diagnoses failures before repairing them, achieves higher yield and recovery rate than naive regeneration. Downstream fine-tuning quality is driven primarily by generator scale, with filtering and recovery providing their clearest gains when smaller generators are used. We release our evaluation corpus and experimental configurations to support reproducible pipeline design.

Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor probability mass into the target model’s vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the mixing strength can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on structured, front-loaded adversarial benchmarks, with a 4.6-6.6 x inference slowdown inside the band of existing inference-time defenses. The results show that early-token refusal behavior can be transferred between model families at inference time, without touching either model’s weights.

Methods for repairing safety alignment in fine-tuned Large Language Models (LLMs) are evolving quickly, but the literature is fragmented across the repair pipeline: each method is typically tested on its own drifted checkpoint, with its own repair implementation, and judged against its own set of baselines. This makes methods hard to compare or select, and calls for a framework that standardizes drift-induction, repair, and evaluation while remaining flexible enough to support new paradigms and architectures. To address this gap, we introduce SAFETUNE, a source-available library that unifies drift-induction, repair, and evaluation behind a standardized pipeline. SAFETUNE hosts four repair paradigms: weight patching, safety-constrained retraining, gradient-based unlearning, and inference-time steering, with implementations audited and debugged against their original papers. A documented intervention guide suggests a starting repair for a checkpoint’s safety-capability profile, with the full pipeline runnable in four lines of code.

Mental manipulation is a subtle yet pervasive form of abuse in interpersonal communication, making its detection critical for safeguarding potential victims. However, due to manipulation’s nuanced and context-specific nature, identifying manipulative language in complex, multi-turn, and multi-person conversations remains a significant challenge for large language models (LLMs). To address this gap, we introduce the MultiManip dataset, comprising 220 multi-turn, multi-person dialogues balanced between manipulative and non-manipulative interactions, all drawn from reality shows that mimic real-world scenarios. For manipulative interactions, it includes 11 distinct manipulations depicting real-life scenarios. We conduct extensive evaluations of state-of-the-art LLMs, such as GPT-4o and Llama-3.1-8B, employing various prompting strategies. Despite their capabilities, these models often struggle to detect manipulation effectively. To overcome this limitation, we propose SELF-PERCEPT, a novel, two-stage prompting framework inspired by Self-Perception Theory, demonstrating strong performance in detecting multi-person, multi-turn mental manipulation. Our code and data are publicly available at https://github.com/danushkhanna/self-percept .

Agriculture, vital for global sustenance, necessitates innovative solutions due to a lack of organized domain experts, particularly in developing countries where many farmers are impoverished and cannot afford expert consulting. Initiatives like Farmers Helpline play a crucial role in such countries, yet challenges such as high operational costs persist. Automating query resolution can alleviate the burden on traditional call centers, providing farmers with immediate and contextually relevant information.The integration of Agriculture and Artificial Intelligence (AI) offers a transformative opportunity to empower farmers and bridge information gaps.Language models like transformers, the rising stars of AI, possess remarkable language understanding capabilities, making them ideal for addressing information gaps in agriculture.This work explores and demonstrates the transformative potential of Large Language Models (LLMs) in automating query resolution for agricultural farmers, leveraging their expertise in deciphering natural language and understanding context. Using a subset of a vast dataset of real-world farmer queries collected in India, our study focuses on approximately 4 million queries from the state of Tamil Nadu, spanning various sectors, seasonal crops, and query types.

Lay summarization aims to generate summaries of technical articles for non-experts, enabling easy comprehension for a general audience. The technical language used in research often hinders effective communication of scientific knowledge, making it difficult for non-experts to understand. Automatic lay summarization can enhance access to scientific literature, promoting interdisciplinary knowledge sharing and public understanding. This has become especially important for biomedical articles, given the current global need for clear medical information. Large Language Models (LLMs), with their remarkable language understanding capabilities, are ideal for abstractive summarization, helping to make complex information accessible to the public. This paper details our submissions to the BioLaySumm 2024 Shared Task: Lay Summarization of Biomedical Research Articles. We fine-tune and evaluate sequence-to-sequence models like T5 across various training dataset settings and optimization methods such as LoRA for lay summarization. Our submission achieved the 53rd position overall.

This paper describes our submission to Task 10 at SemEval 2023-Explainable Detection of Online Sexism (EDOS), divided into three subtasks. The recent rise in social media platforms has seen an increase in disproportionate levels of sexism experienced by women on social media platforms. This has made detecting and explaining online sexist content more important than ever to make social media safer and more accessible for women. Our approach consists of experimenting and finetuning BERT-based models and using a Majority Voting ensemble model that outperforms individual baseline model scores. Our system achieves a macro F1 score of 0.8392 for Task A, 0.6092 for Task B, and 0.4319 for Task C.

This paper describes our approach to submissions made at Shared Task 2 at BLP Workshop - Sentiment Analysis of Bangla Social Media Posts. Sentiment Analysis is an action research area in the digital age. With the rapid and constant growth of online social media sites and services and the increasing amount of textual data, the application of automatic Sentiment Analysis is on the rise. However, most of the research in this domain is based on the English language. Despite being the world’s sixth most widely spoken language, little work has been done in Bangla. This task aims to promote work on Bangla Sentiment Analysis while identifying the polarity of social media content by determining whether the sentiment expressed in the text is Positive, Negative, or Neutral. Our approach consists of experimenting and finetuning various multilingual and pre-trained BERT-based models on our downstream tasks and using a Majority Voting and Weighted ensemble model that outperforms individual baseline model scores. Our system scored 0.711 for the multiclass classification task and scored 10th place among the participants on the leaderboard for the shared task. Our code is available at https://github.com/ptnv-s/RSM-NLP-BLP-Task2 .