Raymond Tsao

Email · GitHub · LinkedIn · CV

Raymond-Tsao.png

Hi! I’m Raymond, a fifth-year master’s student at UC Berkeley studying computer science and applied mathematics. I’m advised by Andrew Wagenmaker and Sergey Levine in the Robotic AI and Learning Lab@BAIR. Previously, I worked on LLM evaluation at Berkeley’s RISELab, where I co-led the development of the Berkeley Function Calling Leaderboard (BFCL) v4 web search benchmark.

I’m seeking research engineer and machine learning engineer roles in robotics, starting after my graduation in Spring 2027.

Research

My research focuses on efficient RL finetuning of pretrained robot policies. Vision language action (VLA) models provide a promising starting point for general-purpose robotics, but deploying them in new real-world settings often requires further adaptation. I am interested in how RL can enable this adaptation efficiently, through two complementary directions:

  • Pretraining for adaptation: How can we pretrain policies that are easier to finetune with RL?
  • Efficient finetuning: How can we speed up RL improvement of existing pretrained policies?

Publications

  1. Learning Process Rewards via Success Visitation Matching for Efficient RL
    Learning Process Rewards via Success Visitation Matching for Efficient RL
    Raymond Tsao*, Andrew Wagenmaker, Sergey Levine
    Under review
    Paper  |  Website  |  We found that a discriminator trained to classify state-action pairs from successful and failed trajectories can be used as a dense process reward, which leads to significantly faster RL finetuning.
  2. Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning
    Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning
    Andrew Wagenmaker*, Perry Dong, Raymond Tsao, Chelsea Finn, Sergey Levine
    ICML 2026 (Spotlight)
    Paper  |  We show that, instead of fitting the observed demonstrations, training a policy to model the posterior distribution of the demonstrator's behavior given the demonstration dataset leads to a policy that is easier to finetune with RL.
  3. Berkeley Function Calling Leaderboard V4: Web Search
    Berkeley Function Calling Leaderboard V4: Web Search
    Huanzhi Mao*, Raymond Tsao*, Jingzhuo Zhou*, Shishir G. Patil, Joseph E. Gonzalez
    Berkeley Function Calling Leaderboard (BFCL) V4
    Blog  |  Leaderboard  |  Code  |  We build a challenging, multihop web search benchmark to evaluate LLM agents' ability to search for and synthesize information across multiple web sources, and identify common failure modes in state-of-the-art LLMs.