Hi I am Yinggan!

I am a researcher, engineer and self-proclaimed artist. I am working on foundation models for robots at XPENG Robotics.

Research-wise, I am currently working on Large Language Models, especially on its reasoning ability and different properties. I am also broadly interested in the domain of automatic scientific discovery, robotics, reinforcement learning and human-AI interaction.

I work closely with Xin Qiu, Prof. Risto Miikkulainen and other amazing folks from Cognizant AI Lab on various research topcis related to LLM, evolution algorithms and the sciences behind. I also worked with Prof. Di Luo on LLM for Science.

I was with Meitu and CFI. I got my MS in Compute Science from UCLA and my BEng in Computer Engineering from CUHKSZ where I work closely with Prof. Baoxiang Wang.

Research

Evolution Strategy for LLM, sometimes better than RFT

Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning

Xin Qiu, Yulu Gan, Conor F. Hayes, Qiyao Liang, Yinggan Xu, Roberto Dailey, Elliot Meyerson, Babak Hodjat, Risto Miikkulainen
ICML 2026
  • First paper on ES as a powerful alternative to reinforcement learning for LLM post-training. @ArXiv
  • Contributed a 10x faster implementation for LLM ES. @Github · @X(Twitter)

Quantized Evolution Strategies: High-precision Fine-tuning of Quantized LLMs at Low-precision Cost

Yinggan Xu, Risto Miikkulainen, Xin Qiu

  • Yes — you can train your quantized model even further at inference cost.

  • QES extends ES to post-training of quantized LLMs. With frugal memory at low-precision inference levels, it achieves a high-precision optimization trajectory in quantized parameter space.

    @Github · @ArXiv · @Linkedin

AI for Science / LLM Agent

PhySense: Principle-Based Physics Reasoning Benchmarking for Large Language Models

Yinggan Xu*, Yue Liu*, Zhiqiang Gao, Changnan Peng, Di Luo
NeurIPS 2025 AI4S Workshop
  • 500+ physics problems evaluating LLMs’ capability of applying physics principles and bypassing complex & costly reasoning.
  • Shorter version at @NeurIPS2025 AI4S Workshop.

Advancing AI-Scientist Understanding: Multi-Agent LLMs with Interpretable Physics Reasoning

Yinggan Xu, Hana Kimlee, Yijia Xiao, Di Luo
ICML 2025 MAS Workshop
  • Multi-agent LLM framework for interpretable physics reasoning.

Reinforcement Learning and Robot Learning

Interactive Imitation Learning

Yinggan Xu

Provably Efficient Convergence of Primal-Dual Actor-Critic with Nonlinear Function Approximation

Jing Dong, Li Shen, Yinggan Xu, Baoxiang Wang
AAMAS 2023
  • Convergence guarantees for primal-dual actor-critic methods under nonlinear function approximation.

About this blog

I learned a lot from people’s blog, and I would like to share some of my thoughts in case they can help others. This is the initial thought of maintaining this blog, or personal website.

Also in this era of powerful LLMs, I guess it is good to have some cognitive burden on the human side to make sure my brain is still functioning.

I wrote most of my technical contents in English and literary things in Chinese.