Abinitha Gourabathina, Inkit Padhi, Manish Nagireddy, Subhajit Chaudhury, Prasanna Sattigeri. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Lucas Monteiro Paes, Dennis Wei, Hyo Jin Do, Hendrik Strobelt, Ronny Luss, Amit Dhurandhar, Manish Nagireddy, Karthikeyan Natesan Ramamurthy, Prasanna Sattigeri, Werner Geyer, Soumya Ghosh. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Developing value-aligned agents is a complex undertaking and an ongoing challenge in the field of AI. Indeed, designing Large Language Models (LLMs) that can balance multiple possibly conflicting moral values based on the context is a problem of paramount importance. In this paper, we propose a system that performs contextual value alignment based on contextual aggregation of possible responses. This aggregation is achieved by integrating a subset of possible LLM responses that are best suited to a user's input while taking into account features extracted about the user's moral preferences. The proposed system trained using the Moral Integrity Corpus displays better alignment to human values than state-of-the-art baselines.

LLMs have shown remarkable capabilities, but precisely controlling their response behavior remains challenging. Existing activation steering methods alter LLM behavior indiscriminately, limiting their practical applicability in settings where selective responses are essential, such as content moderation or domain-specific assistants. In this paper, we propose Conditional Activation Steering (CAST), which analyzes LLM activation patterns during inference to selectively apply or withhold activation steering based on the input context. Our method is based on the observation that different categories of prompts activate distinct patterns in the model's hidden states. Using CAST, one can systematically control LLM behavior with rules like "if input is about hate speech or adult content, then refuse" or "if input is not about legal advice, then refuse." This allows for selective modification of responses to specific content while maintaining normal responses to other content, all without requiring weight optimization. We release an open-source implementation of our framework.

Current datasets for unwanted social bias auditing are limited to studying protected demographic features such as race and gender. In this work, we introduce a comprehensive benchmark that is meant to capture the amplification of social bias, via stigmas, in generative language models. Taking inspiration from social science research, we start with a documented list of 93 US-centric stigmas and curate a question-answering (QA) dataset which involves simple social situations. Our benchmark, SocialStigmaQA, contains roughly 10K prompts, with a variety of prompt styles, carefully constructed to systematically test for both social bias and model robustness. We present results for SocialStigmaQA with two open source generative language models and we find that the proportion of socially biased output ranges from 45% to 59% across a variety of decoding strategies and prompting styles. We demonstrate that the deliberate design of the templates in our benchmark (e.g., adding biasing text to the prompt or using different verbs that change the answer that indicates bias) impacts the model tendencies to generate socially biased output. Additionally, through manual evaluation, we discover problematic patterns in the generated chain-of-thought output that range from subtle bias to lack of reasoning. Warning: This paper contains examples of text which are toxic, biased, and potentially harmful.

Large language models (LLMs) have been rapidly adopted, as showcased by ChatGPT's overnight popularity, and are integrated in products used by millions of people every day, such as search engines and productivity suites. Yet the societal impact of LLMs, encompassing both benefits and harms, is not well understood. Inspired by cybersecurity practices, red-teaming is emerging as a technique to uncover model vulnerabilities. Despite increasing attention from industry, academia, and government centered around red-teaming LLMs, such efforts are still limited in the diversity of the red-teaming focus, approaches and participants. Importantly, given that LLMs are becoming ubiquitous, it is imperative that red-teaming efforts are scaled out to include large segments of the research, practitioners and the people whom are directly affected by the deployment of these systems. The goal of this tutorial is two fold. First, we introduce the topic of LLM red-teaming by reviewing the state of the art for red-teaming practices, from participatory events to automatic AI-focused approaches, exposing the gaps in both the techniques and coverage of the targeted harms. Second, we plan to engage the audience in a hands-on and interactive exercise in LLM red-teaming to showcase the ease (or difficulty) of exposing model vulnerabilities, contingent on both the targeted harm and model capabilities. We believe that the KDD community of researchers and practitioners are in a unique position to address the existing gaps in red-teaming approaches, given their longstanding research and practice of extracting knowledge from data.