Abstract

Poster presentations are a key part of academic conferences, and sometimes, a poster session is the only opportunity we get at presenting our papers. It is thus very important to take full advantage of this opportunity and design a poster that not only attracts people’s attention, but also supports an engaging and memorable presentation. However, I have seen my share of low-effort posters (e.g., just copy-pasting texts and tables from the paper) that don’t do justice to the great and exciting research work that they are supposed to represent. As someone passionate (and judgmental) about graphic design, this hurts my heart (and my eyes…). So, in this tutorial, I hope to i) convince you that designing good posters matters, and ii) provide you with various tools and methods to design better posters. I also have stories to share, so I promise it’s going to be fun!

Bio

Yu Lu Liu is a 2nd year PhD student supervised by Prof. Ziang Xiao. Her research interests sit at the intersection between NLP and Human-Computer Interaction, with a focus on NLP evaluation and responsible AI issues that arise from NLP technologies: how do people actually use NLP systems and how can we bring NLP evaluation closer to that? However, the part of her that is most relevant to this tutorial is that she took graphic design for 3 years in high-school, and that she’s pretty good at giving constructive feedback (read: she likes to complain about things).

Abstract

When discussing politics, people often use subtle linguistic strategies to influence how their audience thinks about issues, which can then impact public opinion and policy. For example, anti-immigration activists may frame immigration as a threat to native born citizens’ jobs, describe immigrants with dehumanizing vermin-related metaphors, or even use coded expressions to covertly connect immigration with antisemitic conspiracy theories. In this talk, I will briefly overview my research program at the intersection of NLP, political framing, and implicitly harmful language, with a focus on computational approaches to analyze dog whistle communication and metaphorical dehumanization. Grounded in an ongoing survey of computational research on antisemitism, I will also discuss the urgent need for group-specific resources, explainable models, and socially-grounded evaluation to understand implicit othering both in language analyzed with LLMs and in the models themselves.

Bio

Julia Mendelsohn is an assistant professor at the University of Maryland College of Information and Institute for Advanced Computer Studies. She is also affiliated with the AI Interdisciplinary Institute at Maryland and the Department of Government and Politics. Her research interests include natural language processing, political communication, and computational sociolinguistics. She is especially interested in developing computational models for understanding subtle and covert rhetoric in online political discussions and the impact of such language, particularly on marginalized communities. Julia has published papers at top-tier natural language processing and computational social science venues, including ACL, NAACL, EMNLP, and ICWSM. Prior to joining UMD, Julia was a postdoc at the University of Chicago Data Science Institute. Julia completed her PhD at the University of Michigan School of Information and received a BA in Linguistics and MS in Computer Science from Stanford University.

Abstract

Harnessing the power of LLMs requires a delicate dance between being helpful and harmless. This creates a fundamental tension between two competing challenges: vulnerability to adversarial attacks that elicit unsafe content, and a tendency for overrefusal on benign but sensitive prompts. Current approaches often navigate this dance with safeguard models that completely reject any content that contains unsafe portions. This approach cuts the music entirely-it may exacerbate overrefusals and fails to provide nuanced guidance for queries it refuses. To teach models a more coordinated choreography, we propose WaltzRL, a novel multi-agent reinforcement learning framework that formulates safety alignment as a collaborative, positive-sum game. WaltzRL jointly trains a conversation agent and a feedback agent, where the latter is incentivized to provide useful suggestions that improve the safety and helpfulness of the conversation agent’s responses. At the core of WaltzRL is a Dynamic Improvement Reward (DIR) that evolves over time based on how well the conversation agent incorporates the feedback. At inference time, unsafe or overrefusing responses from the conversation agent are improved rather than discarded. The feedback agent is deployed together with the conversation agent and only engages adaptively when needed, preserving helpfulness and low latency on safe queries. Our experiments, conducted across five diverse datasets, demonstrate that WaltzRL significantly reduces both unsafe responses (e.g., from 39.0% to 4.6% on WildJailbreak) and overrefusals (from 45.3% to 9.9% on OR-Bench) compared to various baselines. By enabling the conversation and feedback agents to co-evolve and adaptively apply feedback, WaltzRL enhances LLM safety without degrading general capabilities, thereby advancing the Pareto front between helpfulness and harmlessness.