Publications
(* denotes equal contribution; † denotes project lead)
The Era of Real-World Human Interaction: RL from User Conversations
arXiv Preprint / 🔍 Invited Talk at Google and Meta TBD Lab / ⭐️ Paper of the Week by Huggingface, DAIR.AI, and 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. RLHI beats RLHF at the user level and enabling personalized, contextual, and continually improving AI assistants.
SPICE: Self-Play In Corpus Environments Improves Reasoning
arXiv Preprint / 📰 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.
AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling
NeurIPS 2025 (Spotlight)
SoMi-ToM: Evaluating Multi-Perspective Theory of Mind in Embodied Social Interactions
NeurIPS D&B 2025
Do VLMs have internal World Models? Towards an Atomic Evaluation
ACL 2025 Findings / ⭐️ Huggingface Daily Papers Top-3
We introduce WM-ABench, a large-scale benchmark to evaluate whether Vision-Language Models possess internal world models by assessing their perception (visual, spatial, temporal, quantitative, and motion) and prediction (mechanistic simulation, transitive inference, compositional inference).
OS-Genesis: Automating GUI Agent Trajectory Construction via Reverse Task Synthesis
ACL 2025 / ⭐️ Huggingface Daily Papers Top-1
OS-Genesis is 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.
Humanity’s Last Exam
Technical Report / 📰 Featured in The New York Times, Reuters, ...
MuMA-ToM: Multi-modal Multi-Agent Theory of Mind
AAAI 2025 (Oral)
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
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.
How Far Are We From AGI?
TMLR 2024 / ICLR 2024 AGI Workshop (Oral)
Neural Amortized Inference for Nested Multi-agent Reasoning
AAAI 2024 / AAAI 2024 Summer Symposium (Oral)
Multi-agent interactions often rely on higher-order social inference, i.e., understanding how others infer oneself. We introduce a neural amortized inference method to accelerate computationally expensive nested multi-agent reasoning within the I-POMDP framework, significantly reducing computational costs while maintaining high accuracy.
Beyond the Binary: Capturing Diverse Preferences With Reward Regularization
NeurIPS 2024 Workshop on Socially Responsible Language Modelling Research
The Cultural Psychology of Large Language Models
Technical Report
We apply cultural psychology scales to ChatGPT to assess its cognitive processing style and value judgments. We find that the model exhibits Eastern-style holistic processing traits while displaying mixed alignment in its cultural values.
Dynamics of RNA Localization to Nuclear Speckles are Connected to Splicing Efficiency
Science Advances 10 (42), eadp7727
We demonstrate that RNA localization dynamics to nuclear speckles are tied to gene expression by influencing splicing efficiency. Specifically, nuclear speckles coordinate both co- and post-transcriptional splicing regulation by facilitating the removal of inefficiently excised introns in transcripts enriched within them.
OpenCompass: A Universal Evaluation Platform for Foundation Models
Open-source Project / ⭐️ 6K Github Stars
OpenCompass is an LLM evaluation platform, supporting a wide range of models over 100+ datasets.
Fast-DiT: Fast Diffusion Models with Transformers
Open-source Project / ⭐️ 900 Github Stars
Fast-DiT improves the efficiency of Diffusion Transformers (DiTs) by incorporating features such as gradient checkpointing, mixed-precision training, and feature pre-extraction. It delivers a 95% speed increase and a 60% reduction in memory usage, and has been integrated into the official DiT implementation.
