Present day LLMs face the challenge of managing affordance-based safety risks—situations where outputs inadvertently facilitate harmful actions due to overlooked logical implications. Traditional safety solutions, such as scalar outcome-based reward models, parameter tuning, or heuristic decoding strategies, lack the granularity and proactive nature needed to reliably detect and intervene during subtle yet crucial reasoning steps. Addressing this fundamental gap, we introduce AURA, an innovative, multi-layered framework centered around Process Reward Models (PRMs), providing comprehensive, step level evaluations across logical coherence and safety-awareness. Our framework seamlessly combines introspective self-critique, fine-grained PRM assessments, and adaptive safety-aware decoding to dynamically and proactively guide models toward safer reasoning trajectories. Empirical evidence clearly demonstrates that this approach significantly surpasses existing methods, significantly improving the logical integrity and affordance-sensitive safety of model outputs. This research represents a pivotal step toward safer, more responsible, and contextually aware AI, setting a new benchmark for alignment-sensitive applications.

Language models aligned for safety often exhibit fragile and imbalanced mechanisms, increasing the chances of producing unsafe content. In addition, editing techniques to incorporate new knowledge can further compromise safety. To tackle these issues, we propose SafeInfer, a context-adaptive, decoding-time safety alignment strategy for generating safe responses to user queries. safeInfer involves two phases: the 'safety amplification' phase, which uses safe demonstration examples to adjust the model’s hidden states and increase the likelihood of safer outputs, and the 'safety-guided decoding' phase, which influences token selection based on safety-optimized distributions to ensure the generated content adheres to ethical guidelines. Further, we introduce HarmEval, a novel benchmark for comprehensive safety evaluations, designed to address potential misuse scenarios in line with the policies of leading AI technology companies.

In this study, we tackle a growing concern around the safety and ethical use of large language models (LLMs). Despite their potential, these models can be tricked into producing harmful or unethical content through various sophisticated methods, including `jailbreaking' techniques and targeted manipulation. Our work zeroes in on a specific issue: to what extent LLMs can be led astray by asking them to generate responses that are instruction-centric such as a pseudocode, a program or a software snippet as opposed to vanilla text. To investigate this question, we introduce TechHazardQA, a dataset containing complex queries which should be answered in both text and instruction-centric formats (e.g., pseudocodes), aimed at identifying triggers for unethical responses. We query a series of LLMs -- Llama-2-13b, Llama-2-7b, Mistral-V2 and Mistral 8X7B -- and ask them to generate both text and instruction-centric responses. For evaluation we report the harmfulness score metric as well as judgements from GPT-4 and humans. Overall, we observe that asking LLMs to produce instruction-centric responses enhances the unethical response generation by 2-38% across the models. As an additional objective, we investigate the impact of model editing using the ROME technique, which further increases the propensity for generating undesirable content. We observe that the propensity to generate unethical content through instruction-centric responses in comparison to text responses increases significantly with a single edit, rising from an average of 18.9% to 56.7% in zero-shot scenarios, from 31.9% to 56.6% in zero-shot CoT, and from 22.8% to 65.7% in few-shot scenarios.

Somnath Banerjee, Sayan Layek, Hari Shrawgi, Rajarshi Mandal, Avik Halder, Shanu Kumar, Sagnik Basu, Parag Agrawal, Rima Hazra, Animesh Mukherjee. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Recently, influence functions present an apparatus for achieving explainability for deep neural models by quantifying the perturbation of individual train instances that might impact a test prediction. Our objectives in this paper are twofold. First we incorporate influence functions as a feedback into the model to improve its performance. Second, in a dataset extension exercise, using influence functions to automatically identify data points that have been initially ‘silver’ annotated by some existing method and need to be cross-checked (and corrected) by annotators to improve the model performance. To meet these objectives, in this paper, we introduce InfFeed, which uses influence functions to compute the influential instances for a target instance. Toward the first objective, we adjust the label of the target instance based on its influencer(s) label. In doing this, InfFeed outperforms the state-of-the-art baselines (including LLMs) by a maximum macro F1-score margin of almost 4% for hate speech classification, 3.5% for stance classification, and 3% for irony and 2% for sarcasm detection. Toward the second objective we show that manually re-annotating only those silver annotated data points in the extension set that have a negative influence can immensely improve the model performance bringing it very close to the scenario where all the data points in the extension set have gold labels. This allows for huge reduction of the number of data points that need to be manually annotated since out of the silver annotated extension dataset, the influence function scheme picks up ~1/1000 points that need manual correction.

Ensuring the safe alignment of large language models (LLMs) with human values is critical as they become integral to applications like translation and question answering.Current alignment methods struggle with dynamic user intentions and complex objectives, making models vulnerable to generating harmful content.We propose SAFETY ARITH-METIC, a training-free framework enhancing LLM safety across different scenarios: Base models, Supervised fine-tuned models (SFT), and Edited models.SAFETY ARITH-METIC involves Harm Direction Removal to avoid harmful content and Safety Alignment to promote safe responses.Additionally, we present NOINTENTEDIT, a dataset highlighting edit instances that could compromise model safety if used unintentionally.Our experiments show that SAFETY ARITHMETIC significantly improves safety measures, reduces over-safety, and maintains model utility, outperforming existing methods in ensuring safe content generation.Source codes and dataset can be accessed at: https://github.com/ declare

Community Question Answering (CQA) in different domains is growing at a large scale because of the availability of several platforms and huge shareable information among users. With the rapid growth of such online platforms, a massive amount of archived data makes it difficult for moderators to retrieve possible duplicates for a new question and identify and confirm existing question pairs as duplicates at the right time. This problem is even more critical in CQAs corresponding to large software systems like askubuntu where moderators need to be experts to comprehend something as a duplicate. Note that the prime challenge in such CQA platforms is that the moderators are themselves experts and are therefore usually extremely busy with their time being extraordinarily expensive. To facilitate the task of the moderators, in this work, we have tackled two significant issues for the askubuntu CQA platform: (1) retrieval of duplicate questions given a new question and (2) duplicate question confirmation time prediction. In the first task, we focus on retrieving duplicate questions from a question pool for a particular newly posted question. In the second task, we solve a regression problem to rank a pair of questions that could potentially take a long time to get confirmed as duplicates. For duplicate question retrieval, we propose a Siamese neural network based approach by exploiting both text and network-based features, which outperforms several state-of-the-art baseline techniques. Our method outperforms DupPredictor [33] and DUPE [1] by 5% and 7% respectively. For duplicate confirmation time prediction, we have used both the standard machine learning models and neural network along with the text and graph-based features. We obtain Spearman's rank correlation of 0.20 and 0.213 (statistically significant) for text and graph based features respectively. We shall place all our codes and data in the public domain upon acceptance.

Social media platforms were conceived to act as online town squares' where people could get together, share information and communicate with each other peacefully. However, harmful content borne out of bad actors are constantly plaguing these platforms slowly converting them intomosh pits' where the bad actors take the liberty to extensively abuse various marginalised groups. Accurate and timely detection of abusive content on social media platforms is therefore very important for facilitating safe interactions between users. However, due to the small scale and sparse linguistic coverage of Indic abusive speech datasets, development of such algorithms for Indic social media users (one-sixth of global population) is severely impeded.To facilitate and encourage research in this important direction, we contribute for the first time MACD - a large-scale (150K), human-annotated, multilingual (5 languages), balanced (49\% abusive content) and diverse (70K users) abuse detection dataset of user comments, sourced from a popular social media platform - ShareChat. We also release AbuseXLMR, an abusive content detection model pretrained on large number of social media comments in 15+ Indic languages which outperforms XLM-R and MuRIL on multiple Indic datasets. Along with the annotations, we also release the mapping between comment, post and user id's to facilitate modelling the relationship between them. We share competitive monolingual, cross-lingual and few-shot baselines so that MACD can be used as a dataset benchmark for future research.