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

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.

Ethics and Responsible AI Practices

My research group explores the ethical implications of large language models (LLMs) and human-centered AI interactions. A pivotal paper, [HAICOSYSTEM: An Ecosystem for Sandboxing Safety Risks in Human-AI Interactions](http://arxiv.org/abs/2409.16427), presents a framework for assessing the safety risks inherent in these interactions. Additionally, the study [Let Them Down Easy! Contextual Effects of LLM Guardrails on User Perceptions and Preferences](https://arxiv.org/abs/2506.00195) examines how user experiences can be shaped through the implementation of guardrails, emphasizing the responsibility of AI developers. Our work continuously addresses biases and the moral dimensions of deploying AI, as evidenced by the insights from [Mind the Gesture: Evaluating AI Sensitivity to Culturally Offensive Non-Verbal Gestures](https://arxiv.org/abs/2502.17710).

Narrative Analysis and Interpretation

My research group explores the intricacies of narrative analyses, looking at how narratives are constructed and perceived in social contexts. A significant contribution in this area is the paper [Social Story Frames: Contextual Reasoning about Narrative Intent and Reception](https://arxiv.org/abs/2512.15925), which discusses how understanding the intent behind narratives can influence their reception by audiences. Another relevant study, [HEART-felt Narratives: Tracing Empathy and Narrative Style in Personal Stories with LLMs](https://arxiv.org/abs/2405.17633), reveals how narrative style can evoke empathy in readers. We also investigate the interconnectedness of communication styles through [Words Like Knives: Backstory-Personalized Modeling and Detection of Violent Communication](https://arxiv.org/abs/2505.21451), showcasing the importance of context in narrative impact.

Social Intelligence in AI Agents

My research group explores the development of AI agents equipped with social intelligence that can understand and navigate human interactions effectively. A notable contribution is the paper [SoMi-ToM: Evaluating Multi-Perspective Theory of Mind in Embodied Social Interactions](https://arxiv.org/abs/2506.23046), which assesses an AI's ability to comprehend diverse perspectives in social contexts. Furthermore, our work on [SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents](https://arxiv.org/abs/2310.11667) provides a framework to evaluate the social capabilities of AI agents interactively, enhancing their utility in real-world scenarios. Through our studies, we aim to create AI agents that not only assist but also engage meaningfully with humans in social environments.

Innovations in Language Model Adaptation

My research group explores novel techniques for adapting large language models (LLMs) to enhance their relevance and effectiveness in varied applications. A recent paper, [Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoning](https://arxiv.org/abs/2512.02914), presents a new unsupervised metric to gauge the rational capabilities of LLMs, highlighting the importance of reasoning. Another critical advancement is detailed in [ALFA: Aligning LLMs to Ask Good Questions: A Case Study in Clinical Reasoning](https://arxiv.org/abs/2502.14860), which investigates alignment techniques to improve LLMs in medical inquiry scenarios. Our research aims to refine LLMs further, making them more contextually aware and capable of nuanced reasoning.

More about me