Publications

We introduce ThoughtTrace, the first large-scale dataset that pairs real-world multi-turn human-AI conversations with users' self-reported thoughts. Our experiments show that these thoughts are hard 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.

We introduce IdeaScientist, which splits research ideation into gap finding, innovation, and report writing, trains each role with reinforcement learning, and draws solutions from analogous problems across fields using the 2.77M-idea Svalbard Idea Vault. On a 27B open model, it beats the strongest open-source autoresearch baseline by 14.0% and even outperforms Claude Code and Codex agents by up to 5.9%.

We introduce Into the Unknown, a benchmark that gives an agent a science paper and an unpublished follow-up direction from scientists at DOE national laboratories, and asks it to write the follow-up paper. Across nine frontier agents, none of the 135 submissions meets all of the scientists’ conditions, and 114 fail our integrity check, often by claiming work that was never done.

We introduce VideoResearcher, a training-free multi-agent framework that, like a human researcher, autonomously designs, tests, and refines tools for long-video understanding. By evolving its tools from its own tool-use trajectories, it achieves state-of-the-art results among self-improving video agents and approaches the human-designed upper bound.

We introduce the AI Observatory, a public measurement platform that aggregates the most comprehensive set of real AI conversation sources under a common taxonomy of 145 features, spanning function, topic, sensitive use, interaction style, multi-turn dynamics, and conversation structure. Across 23,158 conversations and 85,633 turns, we find that observed AI use is highly heterogeneous with no single source that generalizes, and that occupation-focused analyses omit substantial personal, social, cultural, and safety-relevant interaction.

We let the model itself decide when and how to compact its own context, pairing a compaction tool with a lightweight rubric for when to fire and when to hold off. This self-compaction matches or beats fixed-interval summarization while using 30-70% less context budget.

We introduce HyperTrace, a training-free framework that treats online personalization as tracing a user’s latent preferences: it keeps interpretable natural-language hypotheses about short-term intent and long-term preferences and updates them across turns and sessions. On PRISM and PersonaMem-v2, it improves response alignment, preference prediction, and profile consistency over strong online baselines.

We introduce AutoInteract, which learns user models from real interaction logs and turns verified tasks into realistic multi-turn interactions with clarifications, changing requirements, and feedback, while keeping each task’s executable test. Reinforcement learning on this data substantially improves coding agents, especially in multi-turn settings.

We introduce AutoBenchmark, in which an autoresearch agent builds a benchmark end to end and revises it using solver results and verifier critiques, and we study how much human AI scientists help. Unaided agents produce near-saturated benchmarks, while detailed human direction on what and how to build halves solver scores, showing that humans are still needed in benchmark creation.

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 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.

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.

Humanity's Last Exam (HLE) is a multi-modal benchmark designed to be the final closed-ended academic benchmark with broad subject coverage. I contributed several challenging mathematics and cognitive science questions.

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.

SoMi-ToM is a Minecraft-based benchmark designed to evaluate Vision-Language Models' Theory of Mind capabilities through complex, multi-agent social interactions. Experiments reveal that current models significantly underperform compared to humans.

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.

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).

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.

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.

We explore the trajectory toward Artificial General Intelligence by comprehensively analyzing the requisite capabilities across internal, interface, and system dimensions. We also outline a roadmap for responsibly achieving AGI.

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.

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 is an LLM evaluation platform, supporting a wide range of models over 100+ datasets.

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.