About

I am a first-year PhD student at Tsinghua University and a research intern at Tencent Hunyuan. I work on embodied agents driven by frontier models.

I don't keep count of my papers because they are an old currency. On this page they show up as evidence for a judgment I made at the time. What I care about is calling something early and putting real work behind the call, and when my call turns out wrong it stays here, struck through. I change topics when the technology moves, and I won't retell that as one long line of research, which is the polite fiction of our field. I believe that with the right people and the right taste, technology moves faster than any career plan can follow.

To go deep on any of this, whether to collaborate, consult or just chat, email me anytime.

“One finds limits by pushing them.”

Herbert A. Simon, whom I look up to most

Selected research topicseach with the bet I made going in

Embodied agents driven by frontier models

Bet, Oct 2026Embodied AI should be System 2-centric, with System 1 built around what System 2 needs, rather than the opposite.

JustRUA score progression on RoboDojo

JustRUA: A Controlled Study of GPT-6 Robot-Use Agents

arXiv, Nov 2026Project pageCode

One persistent GPT-6 conversation controls a dual-arm robot in RoboDojo. Controlled ablations show which parts of the harness actually move the score. I led the project.

Reinforcement learning for reasoning

Bet, Dec 2024RL is the key to reasoning. A clever method is a pleasure to build, but the results rest on verifiable rewards and a simple, reliable training stack.

PRIME benchmark results

Process Reinforcement through Implicit Rewards

PRIME, arXiv, Feb 2025Code

Online RL for LLM reasoning with dense token-level rewards, taken from an implicit process reward model that needs outcome labels only. As a core contributor I built the training stack on verl and upstreamed the first parallel verification-reward code in the verl repo.

AutoTriton training pipeline

AutoTriton: Automatic Triton Programming with Reinforcement Learning in LLMs

arXiv, Jul 2025Code

An 8B model trained with SFT and then RL to write Triton GPU kernels. I built the sandbox that verifies the generated kernels.

MiniCPM-V 4.5 architecture

MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe

CVPR 2026Code

An 8B multimodal model with a unified 3D resampler for images and video, and a hybrid RL stage for short and long reasoning. I did the reasoning cold start and nearly all of the final RL experiments.

Hobby

Dota 2A sick mental game that brings out the best and the worst in people. Favorite heroes: Axe and Necrophos. Favorite announcer: The Stanley Parable.

Strategy gamesGames that keep my mind fully engaged, such as Slay the Spire, Stellaris, Yu-Gi-Oh!, Go and no-guess Minesweeper. Most are on hold now, because they use the same part of my brain as research.