Maarten Sap

My research group explores how to evaluate and improve the social intelligence of AI systems, especially their ability to read context, manage information, and respond appropriately in multi-party interaction. A key strand of work asks whether language agents truly understand social situations or merely imitate them, as seen in [SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents](https://arxiv.org/abs/2310.11667) and [Clever Hans or Neural Theory of Mind? Stress Testing Social Reasoning in Large Language Models](https://arxiv.org/abs/2305.14763). More recent work such as [Cognitive Chain-of-Thought: Structured Multimodal Reasoning about Social Situations](https://arxiv.org/abs/2507.20409) and [SOTOPIA-ToM: Evaluating Information Management in Multi-Agent Interaction with Theory of Mind](https://arxiv.org/abs/2605.02307) pushes toward richer, more structured assessments of social reasoning and perspective-taking. Across these papers, the field is moving from static benchmarks to interactive and situation-aware evaluations that better capture real pragmatic competence.

My research group explores how to measure the safety of increasingly agentic AI systems, while also understanding how people rely on them in ways that can help or harm. Work such as [OpenAgentSafety: A Comprehensive Framework for Evaluating Real-World AI Agent Safety](https://arxiv.org/abs/2507.06134) and [HAICOSYSTEM: An Ecosystem for Sandboxing Safety Risks in Human-AI Interactions](http://arxiv.org/abs/2409.16427) reflects a push to test agents in realistic, high-stakes settings rather than only in abstract benchmarks. On the human side, [Rel-A.I.: An Interaction-Centered Approach To Measuring Human-LM Reliance](https://aclanthology.org/2025.naacl-long.556/) and [Relying on the Unreliable: The Impact of Language Models' Reluctance to Express Uncertainty](https://arxiv.org/abs/2401.06730) examine how model behavior shapes trust, dependence, and decision-making. Together, these papers show growing concern with both direct agent failure modes and subtler safety risks such as overreliance, manipulation, and miscalibrated confidence.

My research group explores how AI systems can become more culturally competent, while avoiding harms that arise when personalization or localization ignores community norms. A central theme is benchmarking adaptation across cultures, with [CCBENCH: Assessing LLM Cultural Competence via Implicitly Signaled Norms using Health Queries](https://arxiv.org/abs/2607.05405) and [NormAd: A Framework for Measuring the Cultural Adaptability of Large Language Models](https://aclanthology.org/2025.naacl-long.120/) providing explicit measures of cross-cultural responsiveness. At the same time, fairness work such as [Black LLMirror: User (Self) Perceptions in Black American English Interactions with LLMs](https://dl.acm.org/doi/abs/10.1145/3772318.3791111) highlights how insufficient adaptation can produce alienation, misrecognition, or bias toward minority varieties. This line of research treats cultural competence not as a cosmetic feature, but as a safety and equity requirement for personalized AI.

My research group explores how AI can support human-human connection by understanding stories, personal narratives, and the social meaning embedded in them. [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) study how narrative style and emotional resonance can be identified and modeled computationally. More recent work like [Social Story Frames: Contextual Reasoning about Narrative Intent and Reception](https://arxiv.org/abs/2512.15925) extends this toward understanding how stories are intended, interpreted, and socially received. Overall, these papers suggest that story understanding is becoming a pathway for better empathy modeling, interpersonal communication, and AI systems that help people connect rather than merely generate text.