Research
* Equal contribution.
Building an open index of real-world research tasks that evaluates whether AI agents can drive recursive self-improvement and advance scientific discovery, at scales from a single node to thousands of GPUs.
Project initiator and co-lead
Developed a benchmark of 130 agentic tasks across 35 occupations, evaluating workflows that human experts prioritize for delegation and grading outputs with task-specific rubrics.
Adopted by Meta Muse Spark, Kimi K3, and Qwen3.8
Introduced an expert-grounded benchmark for visual aesthetics across art, photography, and illustration, revealing a substantial gap between 20 frontier vision-language models and expert judgment.
Studied whether automated research agents can produce convincing but unsound papers that deceive LLM reviewers, and evaluated weaknesses in automated peer review.
Best Paper Award, Agents4Science 2025
Built a pipeline that turns open-source repositories and real web pages into 15K verifiable front-end development tasks with verified agent trajectories; fine-tuning Qwen3.5-35B-A3B on them raises WebGen-Bench accuracy from 34.5% to 63.9%.
Introduced a contract in which engineering agents choose relational constraints and specialized tools derive exact consequences, raising the PCB clean-pass rate across nine models to 90.9% from 74.0% for the best baseline.
Proposed a transferable adversarial patch, optimized with only a public visual encoder, that cuts world action model FastWAM's success from over 90% to 0% without querying the target policy.
Characterized how changes in active polyhedral regions govern regret in online learning, with matching theoretical bounds and experiments on combinatorial decision problems.
Established a strict expressivity hierarchy for hypergraph neural networks, identifying structural limitations of bounded-width architectures and motivating density-aware models.
Created large-scale synthetic visual logic puzzles using a rule-to-image pipeline. Applied reinforcement learning to improve vision-language models' logical reasoning and transfer across reasoning tasks.