Maarten Sap

My research group explores how people perceive, trust, and are affected by AI systems in real interactions, with a focus on safety, usability, and responsible deployment. Recent work shows that user-centered framing can meaningfully change reliance and preferences, as in [Framing an AI with Values Reduces AI Reliance in AI-supported Writing Tasks](https://arxiv.org/abs/2605.20512) and [Rel-A.I.: An Interaction-Centered Approach To Measuring Human-LM Reliance](https://aclanthology.org/2025.naacl-long.556/). We also see stronger attention to vulnerable users and high-stakes settings through [Lost in Delusion: Examining LLM Safety Under User Delusions and Distress](https://arxiv.org/abs/2606.00975) and [Why (not) use AI? Analyzing People's Reasoning and Conditions for AI Acceptability](https://arxiv.org/abs/2502.07287). Together, these papers suggest a shift from abstract model evaluation toward measuring what people actually experience, need, and risk in practice.

My research group explores how LLMs and language technologies interpret, generate, and analyze stories, personal narratives, and the social meaning embedded in them. A central thread is narrative intent and reception, highlighted by [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). Related work on [Modeling Empathic Similarity in Personal Narratives](https://arxiv.org/abs/2305.14246) and [The Empirical Variability of Narrative Perceptions of Social Media Texts](https://aclanthology.org/2024.emnlp-main.1113/) shows that narrative interpretation is subjective, context-sensitive, and shaped by reader perspective. This line of research is increasingly useful for understanding how systems process lived experience, persuasion, empathy, and backstory-dependent language use.

My research group explores AI agents that can model social behavior, reason about other minds, and participate in believable multi-party interactions. Recent papers such as [OdysSim: Building Foundation Models for Human Behavior Simulation](https://arxiv.org/abs/2606.14199), [When Should AI Read the Room? Public Perceptions of Social Intelligence in AI Agents](https://arxiv.org/abs/2605.29938), and [Reinforcing Human Behavior Simulation via Verbal Feedback](https://arxiv.org/abs/2605.20506) indicate rapid progress toward agents that can simulate and adapt to human social norms. At the same time, work like [Social World Models](https://arxiv.org/abs/2509.00559) and [SOTOPIA-ToM: Evaluating Information Management in Multi-Agent Interaction with Theory of Mind](https://arxiv.org/abs/2605.02307) emphasizes rigorous evaluation of social reasoning, information management, and theory of mind. Overall, the field is moving from isolated dialogue competence toward richer social simulation and agentic coordination.

My research group explores how AI systems encode values, reproduce cultural bias, and respond to ethically sensitive or socially diverse contexts. A recent cluster of work includes [NormAd: A Framework for Measuring the Cultural Adaptability of Large Language Models](https://aclanthology.org/2025.naacl-long.120/), [EVALUESTEER: Measuring Reward Model Steerability Towards Values and Preference](https://arxiv.org/abs/2510.06370), and [Rejected Dialects: Biases Against African American Language in Reward Models](https://arxiv.org/abs/2502.12858), which together point to the need for models that are both norm-aware and fair across communities. The field is also grappling with social harm and perspective-taking through [PluriHarms: Benchmarking the Full Spectrum of Human Judgments on AI Harm](https://arxiv.org/abs/2601.08951) and [Not Like Us, Hunty: Measuring Perceptions and Behavioral Effects of Minoritized Anthropomorphic Cues in LLMs](https://arxiv.org/abs/2505.05660). These studies suggest that responsible AI now requires measuring harm, adapting to culture, and explicitly accounting for values and identity in model behavior.