My research spans post-training for reasoning and agents, self-evolving agentic systems, and agentic systems for robotics. I study how training signals and execution feedback can improve learning, adaptation, and transfer across tasks and environments. I have also worked on AI safety, including watermarking for language models and synthetic data.
I study learning and adaptation through model post-training, agent harnesses, and robot interaction.
Post-training for Reasoning and Agents
Reinforcement learning, on-policy distillation, and credit assignment for language model reasoning and agents. I study how training objectives, fine-grained signals, and context distributions shape learning efficiency and generalization in reasoning and interactive tasks.
Self-evolving Agentic Systems
Learnable harness evolution for recursive self-improvement, with updates to agent memory, tools, and orchestration learned from execution feedback. I study whether representing harness components as latent states helps these updates generalize across tasks and environments.
Agentic Systems for Robotics
Coding agents that write and execute programs to control robots through device interfaces for zero-shot task execution. I study how composable interfaces, reusable tools, and execution feedback support perception, planning, error recovery, and transfer across tasks and robot embodiments.
A framework for prompt reweighting in reinforcement learning with verifiable rewards. CurveRL uses the rank and density of prompt pass rates to account for their distribution during training.
NeurIPS 2026 (Poster) · Language models · Watermarking
An on-policy fine-tuning framework for watermarking open-weight language models. It uses a watermark signal as a reward while regularizing text quality to improve the quality–detectability trade-off.
Yizhou Zhao, Xiang Li, Peter X. K. Song, Qi Long, Weijie Su
Statistical Learning and Data Science · Synthetic data · Watermarking
A post-editing watermark for synthetic tabular data that embeds signals in the frequency domain. The method handles heterogeneous features and uses rank-based pseudorandom bits for robustness to common data transformations.
EngDesign evaluates language models on practical design tasks across nine engineering domains. Simulation-based evaluation tests whether generated designs meet functional goals and constraints, beyond factual recall or textbook questions.