Maarten Sap

My research group explores how to design, evaluate, and govern AI systems so they better align with human values and everyday needs. Recent work shows that explicit value framing can reduce over-reliance in writing support, while careful privacy redaction explanations can change how people understand AI-mediated interactions: [Framing an AI with Values Reduces AI Reliance in AI-supported Writing Tasks](https://arxiv.org/abs/2605.20512) and [Examining the Effect of Explanations of AI Privacy Redaction in AI-mediated Interactions](https://arxiv.org/abs/2603.24735). We also study how safeguards shape user trust and preferences, as in [Let Them Down Easy! Contextual Effects of LLM Guardrails on User Perceptions and Preferences](https://arxiv.org/abs/2506.00195). Another thread asks how to measure human reliance and preference in a more interaction-centered way, reflected in [Rel-A.I.: An Interaction-Centered Approach To Measuring Human-LM Reliance](https://aclanthology.org/2025.naacl-long.556/). Together, these papers point to a broader agenda of making AI systems more transparent, safer, and more responsive to human judgment.

My research group explores how AI systems interpret, model, and generate stories, especially when narrative context shapes meaning and reception. A central focus is how narrative intent and audience reaction can be represented computationally, highlighted by [Social Story Frames: Contextual Reasoning about Narrative Intent and Reception](https://arxiv.org/abs/2512.15925). We also examine how personal backstory affects violence detection and interpretation in language, as in [Words Like Knives: Backstory-Personalized Modeling and Detection of Violent Communication](https://arxiv.org/abs/2505.21451). Earlier work on [HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs](https://arxiv.org/abs/2405.17633) shows how models can trace empathy and stylistic patterns in personal storytelling. More broadly, this line of research suggests that narrative understanding requires attention to context, perspective, and emotional nuance rather than surface text alone.

My research group explores how AI agents reason about people, coordinate in multi-agent settings, and simulate social interaction. A major recent direction is theory-of-mind style evaluation, exemplified by [SOTOPIA-ToM: Evaluating Information Management in Multi-Agent Interaction with Theory of Mind](https://arxiv.org/abs/2605.02307), which probes how agents manage beliefs and information across interactions. We also study whether human behavior can be simulated and improved in agentic settings, as in [Reinforcing Human Behavior Simulation via Verbal Feedback](https://arxiv.org/abs/2605.20506). Another important thread evaluates the gap between synthetic interaction and real social behavior, emphasized by [Mind the Sim2Real Gap in User Simulation for Agentic Tasks](https://arxiv.org/abs/2603.11245). These papers collectively show that building socially capable agents requires both stronger social reasoning and careful validation against real user behavior.

My research group explores how language models behave across dialects, registers, and languages, with an emphasis on robustness and fairness. Recent work shows that systems can fail under linguistic variation, as in [Out of Style: RAG's Fragility to Linguistic Variation](https://arxiv.org/abs/2504.08231). We also study multilingual moderation and safety, reflected in [PolyGuard: A Multilingual Safety Moderation Tool for 17 Languages](https://arxiv.org/abs/2504.04377). Bias in reward models and ranking systems is another concern, highlighted by [Rejected Dialects: Biases Against African American Language in Reward Models](https://arxiv.org/abs/2502.12858). Together, these papers point toward a research program that treats linguistic diversity as a core requirement for reliable, equitable AI rather than an edge case.