Tianle Li

firstlast@berkeley.edu
I am a Member of Technical Staff at xAI, working on reasoning, post-training, and RL.
- Grok 4 Fast: Co-creator, lead post-training RL training and recipe studies, co-lead distillation and evals.
- Grok 4: Core contributor, synthetic datasets, evals, model tool use.
Previously, I was an EECS undergraduate at UC Berkeley (2021-2025), where I was fortunated to be advised by Ion Stoica and built LMArena. During my undergrad, I also spent a year full-time at Nexusflow as part of the LLM post-training team, working with Banghua Zhu and Jiantao Jiao. Additionally, I briefly interned as a student researcher at Google AI Research, where I was working on reasoning.
I am very interested in fundamental problems in training large models, building more capable and reliable models, and solving superintelligence. Some of the concrete problems I have been thinking about recently:
- How can we transfer intelligence learned in verifiable domains to open-ended questions?
- Design and formulate less hackable and more interpretable reward signals.
- How can we leverage multi-agents to train smarter models.
- How to train faster, more efficient reasoning models / agents?