Maarten Sap

My research group explores how to evaluate and improve the social intelligence of AI systems, with particular attention to pragmatics, intent understanding, and theory-of-mind behavior in realistic interaction settings. Recent work such as [XYBench: Can LLMs Respond Pragmatically to Queries with Misconceptions?](https://arxiv.org/abs/2609.06842) and [SOTOPIA-ToM: Evaluating Privacy and Information Management in Multi-Agent Interaction with Theory of Mind](https://arxiv.org/abs/2605.02307) shows growing interest in whether models can reason about what people mean, not just what they say. [Social Gym and SPaRTan: Benchmarking and Improving LLM Social Reasoning via Multi-Agent Game Tournaments](https://arxiv.org/abs/2608.09128) further reflects a shift toward interactive, competitive settings that stress-test social reasoning under pressure. Together, these papers suggest the field is moving from static benchmarks toward dynamic evaluations of conversation, coordination, and social inference.

My research group explores novel ways to measure and improve the safety of agentic AI systems, while also studying user-centric risks such as over-reliance, manipulation, and harmful interactions. [OpenAgentSafety: A Comprehensive Framework for Evaluating Real-World AI Agent Safety](https://arxiv.org/abs/2507.06134) and [AI-LieDar: Examine the Trade-off Between Utility and Truthfulness in LLM Agents](https://aclanthology.org/2025.naacl-long.595/) highlight concern with how autonomous systems behave when acting on behalf of users. On the human side, [Rel-A.I.: An Interaction-Centered Approach To Measuring Human-LM Reliance](https://aclanthology.org/2025.naacl-long.556/) and [The Hidden Puppet Master: Predicting Human Belief Change in Manipulative LLM Dialogues](https://arxiv.org/abs/2603.20907) focus on when AI influences users too much, or in the wrong direction. This theme points to an emerging safety agenda that treats both agent behavior and human response as part of the same risk surface.

My research group explores how to build AI systems that adapt responsibly across cultures, dialects, and value systems without amplifying bias or exclusion. [NormAd: A Framework for Measuring the Cultural Adaptability of Large Language Models](https://aclanthology.org/2025.naacl-long.120/) and [CCBENCH: Assessing LLM Cultural Competence via Implicitly Signaled Norms using Health Queries](https://arxiv.org/abs/2607.05405) point toward more systematic evaluation of cultural competence in everyday contexts. [Black LLMirror: User (Self) Perceptions in Black American English Interactions with LLMs](https://dl.acm.org/doi/abs/10.1145/3772318.3791111) and [Rejected Dialects: Biases Against African American Language in Reward Models](https://arxiv.org/abs/2502.12858) show how lack of cultural adaptation can translate into fairness harms for minoritized language communities. Overall, the work suggests a shift from treating personalization as a convenience feature to treating it as a core requirement for equitable AI.

My research group explores how AI can support human-human connection by understanding stories, personal narratives, and the social meanings people attach to them. [Social Story Frames: Contextual Reasoning about Narrative Intent and Reception](https://arxiv.org/abs/2512.15925) and [HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs](https://arxiv.org/abs/2405.17633) show how models are being used to analyze not just content, but how stories are told and received. [Modeling Empathic Similarity in Personal Narratives](https://arxiv.org/abs/2305.14246) extends this direction by asking how narrative structure can capture empathy between people. Taken together, these papers suggest a growing effort to make AI better at interpreting stories in ways that preserve nuance, context, and interpersonal meaning.