Maarten Sap

My research group explores how to evaluate and improve the social intelligence of AI systems, especially their ability to understand intent, manage information, and respond pragmatically in realistic interactions. Recent work like [XYBench: Can LLMs Respond Pragmatically to Queries with Misconceptions?](https://arxiv.org/abs/2609.06842) and [Social Gym and SPaRTan: Benchmarking and Improving LLM Social Reasoning via Multi-Agent Game Tournaments](https://arxiv.org/abs/2608.09128) shows a shift toward harder, interaction-based tests rather than static benchmarks. We also see stronger emphasis on mental-state modeling and privacy-aware social reasoning through [SOTOPIA-ToM: Evaluating Privacy and Information Management in Multi-Agent Interaction with Theory of Mind](https://arxiv.org/abs/2605.02307) and [TOM-SWE: User Mental Modeling For Software Engineering Agents](https://arxiv.org/abs/2510.21903). Overall, this theme is moving from whether models can produce socially fluent text to whether they can sustain nuanced, cooperative, and context-sensitive behavior across real interactions.

My research group explores novel measures of agentic AI safety and user-centric safety, spanning autonomous-agent risks, deceptive behavior, overreliance, and safety under distressed or vulnerable use. A major thread is agent evaluation in the wild, highlighted by [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/). On the human side, [Rel-A.I.: An Interaction-Centered Approach To Measuring Human-LM Reliance](https://aclanthology.org/2025.naacl-long.556/) and [Lost in Delusion: Examining LLM Safety Under User Delusions and Distress](https://arxiv.org/abs/2606.00975) reflect growing concern with how AI systems shape belief, dependence, and harm in sensitive contexts. The broader direction is toward safety metrics that capture both what agents do and how people respond to them.

My research group explores responsible AI cultural competence and adaptability, including how systems respond across global contexts, dialects, and identity-linked language variation. [NormAd: A Framework for Measuring the Cultural Adaptability of Large Language Models](https://aclanthology.org/2025.naacl-long.120/) and [NormViz: A Benchmark and Framework for Grounding Multimodal Reasoning in Global Cultures](https://openreview.net/forum?id=nfWUdg2qOE) show a push toward explicit evaluation of cross-cultural understanding rather than assuming one-size-fits-all behavior. We also see important work on fairness and representation in language use through [CCBENCH: Assessing LLM Cultural Competence via Implicitly Signaled Norms using Health Queries](https://arxiv.org/abs/2607.05405) and [Rejected Dialects: Biases Against African American Language in Reward Models](https://arxiv.org/abs/2502.12858). Together, these papers suggest a field increasingly focused on preventing cultural mismatch, dialect-based bias, and personalization that fails to respect local norms.

My research group explores how AI can support human-human connection by understanding stories, narrative framing, and the social meaning of personal experience. A central development is [Social Story Frames: Contextual Reasoning about Narrative Intent and Reception](https://arxiv.org/abs/2512.15925), which points to story interpretation as a social reasoning problem rather than just content analysis. Complementing that, [HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs](https://arxiv.org/abs/2405.17633) and [Modeling Empathic Similarity in Personal Narratives](https://arxiv.org/abs/2305.14246) examine how narrative style and emotional resonance can be modeled computationally. This theme suggests growing interest in AI systems that can better read, compare, and respond to stories in ways that preserve empathy and interpersonal meaning.