Maarten Sap

My research group explores how to evaluate and improve the social intelligence of AI systems, especially whether they can read context, infer intent, and respond pragmatically in multi-turn interaction. A key thread is benchmarking social reasoning and theory of mind with tasks that go beyond surface fluency, such as [SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents](https://arxiv.org/abs/2310.11667) and [FANToM: A Benchmark for Stress-testing Machine Theory of Mind in Interactions](https://arxiv.org/abs/2310.15421). More recent work like [Social Gym and SPaRTan: Benchmarking and Improving LLM Social Reasoning via Multi-Agent Game Tournaments](https://arxiv.org/abs/2608.09128) and [XYBench: Can LLMs Respond Pragmatically to Queries with Misconceptions?](https://aclanthology.org/2026.emnlp-main.2646/) pushes toward dynamic, interaction-based evaluation rather than static test sets. We are also interested in how users perceive these capabilities, as in [When Should AI Read the Room? Public Perceptions of Social Intelligence in AI Agents](https://arxiv.org/abs/2605.29938), which helps connect model performance to real-world expectations. Together, these papers suggest a shift from generic language ability toward socially grounded, situation-aware AI.

My research group explores how to measure and reduce the risks introduced by agentic AI systems and by human over-trust in AI assistance. One major strand studies real-world agent safety and failure modes, exemplified 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/). Another line focuses on human-AI reliance and how system design shapes people’s decisions, including [Rel-A.I.: An Interaction-Centered Approach To Measuring Human-LM Reliance](https://aclanthology.org/2025.naacl-long.556/) and [Right Advice, Wrong Reasons: How Decision and Reasoning Alignment Shape Reliance on AI Advisors](https://arxiv.org/abs/2611.15732). Safety research here also looks at adversarial or vulnerable interaction contexts, such as [Lost in Delusion: Examining LLM Safety Under User Delusions and Distress](https://arxiv.org/abs/2606.00975), where the challenge is not just preventing harm but supporting safe, appropriate interaction under distress. Overall, this theme emphasizes that agentic safety is inseparable from how humans interpret, depend on, and are influenced by AI behavior.

My research group explores how to build AI systems that are culturally competent, adaptable, and fair across diverse linguistic and social contexts. A central focus is measuring whether models can recognize implicit norms and local values, as in [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/). We also study the ethical risks of missing cultural adaptation, including bias against marginalized language varieties, highlighted by [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). More broadly, [Not Like Us, Hunty: Measuring Perceptions and Behavioral Effects of Minoritized Anthropomorphic Cues in LLMs](https://arxiv.org/abs/2505.05660) shows that cultural cues can change user experience and model behavior in subtle ways. The goal is to make personalization and localization responsible rather than merely stylistic.

My research group explores how AI can support human-human connection by understanding stories, narrative intent, and the emotional structure of personal accounts. A key direction is modeling how people interpret and respond to narratives, including [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). More recent work such as [Social Story Frames: Contextual Reasoning about Narrative Intent and Reception](https://arxiv.org/abs/2512.15925) and [How Do People Challenge Racial Stereotypes Online? Counter-Story Detection Across Reddit Communities](https://arxiv.org/abs/2610.04803) extends this toward understanding how stories function in social argument, identity, and community dialogue. These papers suggest that narrative AI is not just about summarization or generation, but about recognizing why stories are told and how they shape empathy, resistance, and shared meaning. In this area, story understanding becomes a tool for better interpersonal and social communication rather than a purely text-level task.