Maarten Sap Twitter Profile Google Scholar Profile Semantic Scholar Profile

Maarten Sap

I am an assistant professor at CMU's LTI department with a courtesy appointment in HCII, and a part-time research scientist and AI safety lead at the Allen Institute for AI (AI2). My research focuses on (1) measuring and improving AI systems' social and interactional intelligence, (2) assessing and combatting social inequality, safety risks, and socio-cultural biases in human- or AI-generated language, and (3) building narrative language technologies for prosocial outcomes. I was named a 2025 Packard Fellow and a recipient of the 2025 Okawa Research Award.

I received my PhD from the University of Washington where I was advised by Noah Smith and Yejin Choi.
[bio for talks]

Recent updates:

December 2025 🏅📃: Very excited to have our paper Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) selected for a Best Paper Award at NeurIPS 2025 (Datasets and Benchmarks Track)!! Huge congrats to the first author Liwei Jiang!!!

November 2025 💎🚀: Honored to be a Spring 2025 recipient of the Amazon Research Award for our project on measuring AI agentic safety!

October 2025 🏅⭐: I’m super excited and grateful to announce that I'm part of the 2025 class of Packard Fellows. The Packard Foundation and this fellowship will allow me to explore exciting research directions towards culturally responsible and safe AI 🌍🌈

October 2025 🔍🧑‍🎓: Due to my lab being quite full already, I'm not taking looking for any new students in this upcoming PhD application cycle 😟.

October 2025 🇨🇦🎉: Excited to be attending COLM 2025 in Montreal this October! I'll be giving a talk at the Social Sim Workshop on Unlocking Social Intelligence in AI agents. I'm also thrilled that five papers I co-authored will be presented by my amazing collaborators at COLM: HAICOSYSTEM: An Ecosystem for Sandboxing Safety Risks in Human-AI Interactions (led by Xuhui Zhou et al.), ALFA: Aligning LLMs to Ask Good Questions: A Case Study in Clinical Reasoning (co-led by Jimin Mun et al.), PolyGuard: A Multilingual Safety Moderation Tool for 17 Languages, Fluid Language Model Benchmarking, and The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains.

August 2025 🌟: Incredibly honored to be one of 7 US recipients of the 2025 Okawa Research Grant from the Okawa Foundation!

August 2025 🧑‍🎓: Welcoming my first postdoc, Vasudha Varadarajan, to the lab!

[older news]

My research group:

Dan Chechelnitsky

CMU Portugal LTI PhD student
co-advised with Chrysoula Zerva

Joel Mire

LTI PhD student

Karina Halevy

LTI PhD student
co-advised with Mona Diab

Malia Morgan

Pre-doctoral Young Investigator at Ai2

Jimin Mun

LTI PhD student

Jocelyn Shen

MIT PhD student
co-advised with Cynthia Breazeal

Kynnedy Smith

HCII PhD student
co-advised with Motahhare Eslami

Vasudha Varadarajan

LTI Postdoc

Akhila Yerukola

LTI PhD student

Mingqian Zheng

LTI PhD student
co-advised with Carolyn Rosé

Xuhui Zhou

LTI PhD student

Overarching Research Themes

Themes extracted and images generated with the OpenAI API; there may be inconsistencies.

Social Reasoning and Pragmatics

My research group explores how to evaluate and improve AI systems' ability to navigate social situations, including theory of mind, non-literal language, and context-sensitive interaction. Recent work such as [Cognitive Chain-of-Thought](https://arxiv.org/abs/2507.20409) shows how structured multimodal reasoning can help models interpret social scenes more reliably, while [SOTOPIA-ToM](https://arxiv.org/abs/2605.02307) stress-tests whether agents can manage information strategically in multi-agent interactions. [Social World Models](https://arxiv.org/abs/2509.00559) and [OdysSim](https://arxiv.org/abs/2606.14199) push toward richer simulations of human behavior, which are useful both for benchmarking and for training better social understanding. Together, this line of work suggests that social intelligence is becoming a core capability that needs dedicated evaluation rather than incidental prompting.

Agentic and User Safety Metrics

My research group explores new ways to measure whether agentic AI systems are genuinely safe, especially when safety depends on human interaction, reliance, or manipulation. [OpenAgentSafety](https://arxiv.org/abs/2507.06134) lays out a broader framework for evaluating risks in real-world agent deployments, while [Rel-A.I.](https://aclanthology.org/2025.naacl-long.556/) focuses on measuring human-LM reliance in interaction-centered settings. [The Hidden Puppet Master](https://arxiv.org/abs/2603.20907) highlights the dangers of persuasive or manipulative dialogue by predicting belief change, and [Useless but Safe?](https://arxiv.org/abs/2604.27093) studies the trade-off between safe clarification and task utility. This research area is moving beyond static harmful-content detection toward measuring how systems behave over time in consequential human workflows.

Culturally Adapted and Fair AI

My research group explores how to make AI systems culturally competent, adaptable, and less biased when interacting across languages, norms, and communities. [CCBENCH](https://arxiv.org/abs/ ) and [NormViz](https://arxiv.org/abs/ ) target cultural norm understanding in settings where implicit expectations and global context matter, while [NormAd](https://aclanthology.org/2025.naacl-long.120/) provides a framework for measuring cultural adaptability more systematically. [Black LLMirror](https://dl.acm.org/doi/abs/10.1145/3772318.3791111) and [Rejected Dialects](https://arxiv.org/abs/2502.12858) reveal how dialect- and identity-linked language can trigger harmful model behavior or unfair reward-model bias. The broader trend is toward treating cultural mismatch as a measurable safety and fairness problem, not just a quality issue.

Story Understanding for Connection

My research group explores how AI can support human-human connection by understanding stories, personal narratives, and the meanings people attach to them. [Social Story Frames](https://arxiv.org/abs/2512.15925) examines contextual reasoning about narrative intent and reception, while [HEART-felt Narratives](https://arxiv.org/abs/2405.17633) studies empathy and narrative style in personal stories. [Modeling Empathic Similarity in Personal Narratives](https://arxiv.org/abs/2305.14246) further shows how models can characterize emotional and interpretive similarity across stories, and [Words Like Knives](https://arxiv.org/abs/2505.21451) connects story backstories to the detection of violent communication. Overall, this work treats narrative understanding as a path toward more empathetic, socially grounded AI that can better mediate and interpret human experience.

More about me