Hello! I am a 2nd-year Ph.D. student in Computer Science at Johns Hopkins University, advised by Tianmin Shu. I’m also a student researcher at Meta FAIR, mentored by Jason Weston. I’m honored to be named an Amazon AI PhD Fellow.
I received bachelor’s degrees in Honors Computer Science and Mathematics from NYU Courant (2020–2024), where the CS Department recognized me as its most promising student. I was as a research intern at MIT (2023), advised by Josh Tenenbaum. I received an Outstanding Paper Award at ACL 2024 for my work on multimodal Theory of Mind.
I’m developing AI models and agents with advanced social intelligence (ASI). Particularly, I am working on:
▶Reasoning about Human Interaction
- I lead MMToM-QA (ACL'24 Outstanding Paper) and AutoToM (NeurIPS'25 Spotlight), pushing the frontier of machine Theory of Mind (ToM)—teaching AI systems to understand people's minds (e.g., goals, beliefs) from their behavior.
- We bridge the deliberate human-like reasoning of agent models with the adaptability of language models, advancing ToM by grounding it in cognitive theories and multi-modal / multi-agent / multi-perspective data, enabling self-supervised learning from human actions, and scaling inference to long-context, complex scenarios.
▶Learning from Human Interaction
- We advocate for an era of real-world human interaction, where models learn from natural human interaction for continual improvement and multifaceted alignment. We introduce Reinforcement Learning from Human Interaction (RLHI) as a starting point.
- We present the ThoughtTrace dataset for understanding and learning from user thoughts in real-world LLM interactions.
- I co-organized Continual Robot Learning from Humans @ RSS 2025 and ReLearn @ CVPR 2026.
News
- Feb 2026Co-organizing CVPR 2026 Workshop on “Rediscovering Intelligence: Can AI Still Learn from Humans?”.
- Oct 2025Invited talks on “Reinforcement Learning from Human Interaction” at Google DeepMind, Meta TBD Lab, WashU, and Northwestern.
- Oct 2025Excited to be awarded the Amazon AI PhD Fellowship!
- Sep 2025AutoToM is accepted by NeurIPS 2025 as a Spotlight Presentation!
- Jun 2025Lead-organized RSS 2025 Workshop on Continual Robot Learning from Humans.
- Jun 2025I was selected as a Notable Reviewer for ICLR 2025.
- Jan 2025MuMA-ToM is accepted by AAAI 2025 as an Oral Presentation!
- Aug 2024MMToM-QA won the Outstanding Paper Award at ACL 2024! It was covered by Futurity, Synced, and JHU News.
Recent Publications
(* denotes equal contribution; † denotes project lead)
For a more complete list, please see my publication page or Google Scholar page.
For a more complete list, please see my publication page or Google Scholar page.
ThoughtTrace: Understanding User Thoughts in Real-World LLM Interactions
Tech Report / ⭐️ Paper of the Week by Huggingface / 📰 Featured in Synced
We introduce ThoughtTrace, the first large-scale dataset that pairs real-world multi-turn human-AI conversations with users' self-reported thoughts. Our analysis shows that these thoughts are distinct from messages, difficult for frontier LLMs to infer, and provide actionable signals for user-behavior prediction and model alignment, opening new directions for user modeling, model training, and evaluation.
The Era of Real-World Human Interaction: RL from User Conversations 
Tech Report / 🔍 Invited Talk at Google DeepMind, Meta TBD Lab / ⭐️ Paper of the Week by Huggingface, DAIR.AI, TuringPost
Tech Report / 🔍 Invited Talk at Google DeepMind, Meta TBD Lab / ⭐️ Paper of the Week by Huggingface, DAIR.AI, TuringPost We posit that to achieve continual model improvement and multifaceted alignment, future models must learn from natural human interaction. We introduce Reinforcement Learning from Human Interaction (RLHI), a paradigm that learns directly from in-the-wild user conversations, leveraging organic replies and long-term history as learning signals. Trained on WildChat, RLHI outperforms RLHF in personalization and instruction-following, and similar feedback enhances performance on reasoning benchmarks.
SPICE: Self-Play In Corpus Environments Improves Reasoning Tech Report / 📰 Featured in VentureBeat
Tech Report / 📰 Featured in VentureBeat SPICE is a reinforcement learning framework where a single model improves itself by playing two roles: a Challenger that creates tasks based on corpora, and a Reasoner that solves them. By grounding this self-play in corpora, SPICE addresses hallucination and lack of diversity issues, significantly outperforming standard (ungrounded) self-play across reasoning benchmarks.
MindZero: Learning Online Mental Reasoning With Zero Annotations
ICML 2026 / HCAIR@ICLR 2026 (Oral)
We introduce MindZero, a self-supervised reinforcement learning framework that trains multimodal LLMs for efficient and robust online Theory-of-Mind reasoning. During training, the model is rewarded for generating mental state hypotheses that best explain observed human actions. Across four experimental settings, MindZero significantly outperforms LLMs and model-based methods in accuracy and efficiency.
AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling
NeurIPS 2025 (Spotlight)
AutoToM is an automated agent modeling method for scalable, robust, and interpretable mental inference. Given a ToM problem, AutoToM proposes a minimal agent model, performs Bayesian inverse planning based on the model, and iteratively refines the model until it can confidently infer the target mental state. It achieves SOTA on five benchmarks, produces human-like confidence estimates, and supports embodied decision-making.
OS-Genesis: Automating GUI Agent Trajectory Construction via Reverse Task Synthesis
ACL 2025 / ⭐️ Huggingface Daily Papers Top-1
We introduce OS-Genesis, a manual-free data pipeline for synthesizing GUI agent trajectory. It enables agents to actively explore web and mobile environments through stepwise interactions, then derive meaningful low- and high-level task instructions from observed interactions and state changes.
MuMA-ToM: Multi-modal Multi-Agent Theory of Mind
AAAI 2025 (Oral) / ⭐️ Featured as a CVPR 2026 Challenge
MuMA-ToM evaluates Theory of Mind reasoning in embodied multi-agent interactions, revealing that current multimodal LLMs significantly lag behind human performance. To bridge this gap, we propose LIMP, a method that combines language models with inverse multi-agent planning to achieve superior results.
MMToM-QA: Multimodal Theory of Mind Question Answering
ACL 2024 (Outstanding Paper Award) / 🔍 Invited Talk at University of Washinton / 📰 Featured in Futurity, Synced, ...
Can machines understand people's minds from multimodal inputs? We introduce a comprehensive benchmark, MMToM-QA, and highlight key limitations in current multimodal LLMs. We then propose a novel method that combines the flexibility of LLMs with the robustness of Bayesian inverse planning, achieving promising results.
My research journey began with these undergrad projects. I’m deeply grateful to the mentors who have guided me along the way and inspired me with their vision and passion.
Selected Honors & Awards
- Amazon AI PhD Fellowship2025
- Notable Reviewer Award, ICLR2025
- Outstanding Paper Award, ACL2024
- Presidential Honors Scholar and Summa cum Laude, New York University2024
- Computer Science Prize for the Most Promising Student, New York University (1 person/year)2023
- Dean's Undergraduate Research Fund, New York University2023
- COMAP International Scholarship Award (Top 0.1%)2022
- MAA Award in Mathematical Contest in Modeling (Top 0.1%)2022
- Bronze Medal of Shing-Tung Yau Computer Science Award (Top 1%)2019
- Finalist of FIRST Robotics Competition World Championship (Top 0.2%)2019
- NFLS Outstanding Student Leader Award and Zhou Enlai Scholarship (Top 1%)2018
- First Prize of Chinese Mathematical Olympiad (Top 0.1%)2018
- Champion of International Regions Mathematics League2018
Services
- Workshop Co-organizer: Rediscovering Intelligence: Can AI Still Learn from Humans? (ReLearn @ CVPR 2026) and Continual Robot Learning from Humans (CRLH @ RSS 2025)
- Conference Program Committee / Reviewer: Machine Learning (ICLR - Notable Reviewer Award, NeurIPS, ICML, AAAI), Natural Language Processing (ARR, ACL, EMNLP, NAACL), Computer Vision (CVPR), Robotics (RSS)
- Head Teaching Assistant: JHU EN.601 Machine Social Intelligence with Tianmin Shu
