Shuyao “Tim” Xu
I am an MSCS student at Stanford University.
My
research explores how AI systems, from LLMs to embodied agents, can
learn
What I mean by efficient learning
Foundation models for language and embodied intelligence are increasingly capable,
but rollouts on difficult, long-horizon tasks remain costly. I am exploring how agents
can reuse prior trajectories more effectively—especially with delayed or offline
feedback—and how they can learn from signals beyond the final reward.
and run
What I mean by running efficiently
Capable agents can be expensive to operate, especially when long contexts, repeated
tool use, and perception-action loops compound inference costs. I am interested in
efficient architectures, inference systems, and adaptive compute that can make these
agents more responsive, affordable, and broadly deployable.
more efficiently. I hope these gains can broaden access to capable AI and
contribute to a future of shared abundance.
I am grateful to be advised and mentored by Prof. Bryan Hooi, Prof. Yu Meng, and Dr. Weidi Xu, and to work with many other amazing collaborators.
I’d love to collaborate on LLMs and embodied agents. Why both Why both? These areas are increasingly converging. Embodied agents are adopting VLA-style paradigms and confronting many questions familiar from LLM research: scaling, long-context memory, and fast and slow reasoning. I hope to carry insights from my LLM work into embodied intelligence, which I believe offers a broad arena for real-world impact. Feel free to reach out if you’d like to chat about anything.
Research
For ongoing and exploratory work, see my research notes .
Experience
TikTok
Jun 2025 – Jan 2026- Improved account search relevance.
INF AI
Dec 2024 – May 2025- Post-trained INFLogic-32B-RL via online RL. SOTA on ZebraLogicBench (85.1%).
Education
