Maarten Sap

My research group explores how to evaluate and improve the social intelligence of AI systems, especially their ability to reason about intentions, misconceptions, and privacy in interaction. 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) show a move toward more demanding, interactive assessments rather than static prompt tests. We are also seeing richer modeling of social situations in [Cognitive Chain-of-Thought: Structured Multimodal Reasoning about Social Situations](https://arxiv.org/abs/2507.20409) and [Social World Models](https://arxiv.org/abs/2509.00559), which aim to represent social context more explicitly. Together, these papers suggest the field is shifting from whether models can answer socially to whether they can behave appropriately in dynamic, human-like settings.

My research group explores new ways to measure and improve the safety of agentic AI, while also studying how people rely on, trust, or are manipulated by these systems. [OpenAgentSafety: A Comprehensive Framework for Evaluating Real-World AI Agent Safety](https://arxiv.org/abs/2507.06134) captures the push toward realistic safety evaluation for agents operating in open-ended environments. 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) highlight concern with overreliance, persuasion, and subtle harms in interaction. Related work such as [AI-LieDar: Examine the Trade-off Between Utility and Truthfulness in LLM Agents](https://aclanthology.org/2025.naacl-long.595/) shows that safety is increasingly being studied as a trade-off between helpfulness, autonomy, and honesty.

My research group explores how to build AI systems that adapt responsibly across cultures, dialects, and value systems without amplifying bias or erasing marginalized users. [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) reflect growing interest in measuring whether models can understand culturally specific norms rather than defaulting to a narrow worldview. At the same time, papers like [Rejected Dialects: Biases Against African American Language in Reward Models](https://arxiv.org/abs/2502.12858) and [Black LLMirror: User (Self) Perceptions in Black American English Interactions with LLMs](https://dl.acm.org/doi/abs/10.1145/3772318.3791111) show how lack of adaptation can create unfair experiences for dialect speakers. Work on [CCBENCH: Assessing LLM Cultural Competence via Implicitly Signaled Norms using Health Queries](https://arxiv.org/abs/2607.05405) underscores that cultural competence is becoming a core evaluation axis, not an afterthought.

My research group explores how AI can support human-human connection by understanding stories, narratives, and the social meaning embedded in them. [Social Story Frames: Contextual Reasoning about Narrative Intent and Reception](https://arxiv.org/abs/2512.15925) points to the importance of modeling not just what a story says, but how it is meant to be received. [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) show how narrative analysis can be used to better capture empathy, shared experience, and interpersonal understanding. This direction suggests that story-aware AI could help people interpret one another more carefully, especially in contexts where meaning depends on lived experience and perspective.