Scholar
Yuexiang Zhai
Google Scholar ID: 78WTKm4AAAAJ
UC Berkeley | Google DeepMind
Citations & Impact
Citations
2,400
H-index
18
i10-index
19
Publications
20
Co-authors
53
list available
Publications
Resume (English only)
Academic Achievements
- Co-first author of 'SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training' at ICML 2025.
- First author of 'Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning' at NeurIPS 2024.
- Co-author of 'Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs', presented orally at CVPR 2024.
- First author of 'Investigating the Catastrophic Forgetting in Multimodal Large Language Model Fine-Tuning' at CPAL 2024.
- Co-first author of 'Cal-QL: Calibrated Offline RL Pre-Training for Efficient Online Fine-Tuning' at NeurIPS 2023.
- Co-first author of 'Understanding the Complexity Gains of Single-Task RL with a Curriculum' at ICML 2023.
- First author of 'Computational Benefits of Intermediate Rewards for Goal-Reaching Policy Learning' published in JAIR 2022.
- Contributed to the Gemini 2.5 project on advanced reasoning, multimodality, long context, and next-generation agentic capabilities.
Background
- Currently a Member of Technical Staff at xAI.
- Research interests span machine learning, reinforcement learning, and large models.
- Curious about things not yet understood, such as how the universe works and what his cats are thinking.
- Recently developed an interest in critiquing 'fake papers'—those that appear fancy but lack real substance.
- Currently focused on near-future (≤5 years) practical directions like data cleaning, evaluation, post-training, and agents, especially in large multimodal models.
