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.

Shanu Kumar, Akhila Yesantarao Venkata, Shubhanshu Khandelwal, Bishal Santra, Parag Agrawal, Manish Gupta. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Chain-of-thought (CoT) prompting has significantly enhanced the the capability of large language models (LLMs) by structuring their reasoning processes. However, existing methods face critical limitations: handcrafted demonstrations require extensive human expertise, while trigger phrases are prone to inaccuracies. In this paper, we propose the Zero-shot Uncertainty-based Selection (ZEUS) method, a novel approach that improves CoT prompting by utilizing uncertainty estimates to select effective demonstrations without needing access to model parameters. Unlike traditional methods, ZEUS offers high sensitivity in distinguishing between helpful and ineffective questions, ensuring more precise and reliable selection. Our extensive evaluation shows that ZEUS consistently outperforms existing CoT strategies across four challenging reasoning benchmarks, demonstrating its robustness and scalability.

Large language models (LLMs) have become integral to various real-world applications, leveraging massive, web-sourced datasets like Common Crawl, C4, and FineWeb for pretraining. While these datasets provide linguistic data essential for high-quality natural language generation, they often contain harmful content, such as hate speech, misinformation, and biased narratives. Training LLMs on such unfiltered data risks perpetuating toxic behaviors, spreading misinformation, and amplifying societal biases which can undermine trust in LLM-driven applications and raise ethical concerns about their use. This paper presents a large-scale analysis of inappropriate content across these datasets, offering a comprehensive taxonomy that categorizes harmful webpages into Topical and Toxic based on their intent. We also introduce a prompt evaluation dataset, a high-accuracy Topical and Toxic Prompt (TTP), and a transformer-based model (HarmFormer) for harmful content filtering. Additionally, we create a new multi-harm open-ended toxicity benchmark (HAVOC) and provide crucial insights into how models respond to adversarial toxic inputs. Our work offers insights into ensuring safer LLM pretraining and serves as a resource for Responsible AI (RAI) compliance. Disclaimer: This paper includes potentially offensive content due to the nature of the research.

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.

Shanu Kumar, Soujanya Abbaraju, Sandipan Dandapat, Sunayana Sitaram, Monojit Choudhury. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. 2023.

Massively Multilingual Transformer based Language Models have been observed to be surprisingly effective on zero-shot transfer across languages, though the performance varies from language to language depending on the pivot language(s) used for fine-tuning.In this work, we build upon some of the existing techniques for predicting the zero-shot performance on a task, by modeling it as a multi-task learning problem.We jointly train predictive models for different tasks which helps us build more accurate predictors for tasks where we have test data in very few languages to measure the actual performance of the model.Our approach also lends us the ability to perform a much more robust feature selection, and identify a common set of features that influence zero-shot performance across a variety of tasks.

In this paper, we aim to solve for unsupervised domain adaptation of classifiers where we have access to label information for the source domain while these are not available for a target domain. While various methods have been proposed for solving these including adversarial discriminator based methods, most approaches have focused on the entire image based domain adaptation. In an image, there would be regions that can be adapted better, for instance, the foreground object may be similar in nature. To obtain such regions, we propose methods that consider the probabilistic certainty estimate of various regions and specific focus on these during classification for adaptation. We observe that just by incorporating the probabilistic certainty of the discriminator while training the classifier, we are able to obtain state of the art results on various datasets as compared against all the recent methods. We provide a thorough empirical analysis of the method by providing ablation analysis, statistical significance test, and visualization of the attention maps and t-SNE embeddings. These evaluations convincingly demonstrate the effectiveness of the proposed approach.

Citizen engagement and technology usage are two emerging trends driven by smart city initiatives. Typically, citizens report issues, such as broken roads, garbage dumps, etc. through web portals and mobile apps, in order for the government authorities to take appropriate actions. Several mediums - text, image, audio, video - are used to report these issues. Through a user study with 13 citizens and 3 authorities, we found that image is the most preferred medium to report civic issues. However, analyzing civic issue related images is challenging for the authorities as it requires manual effort. In this work, given an image, we propose to generate a Civic Issue Graph consisting of a set of objects and the semantic relations between them, which are representative of the underlying civic issue. We also release two multi-modal (text and images) datasets, that can help in further analysis of civic issues from images. We present an approach for adversarial adaptation of existing scene graph models that enables the use of scene graphs for new applications in the absence of any labelled training data. We conduct several experiments to analyze the efficacy of our approach, and using human evaluation, we establish the appropriateness of our model at representing different civic issues.