Yuxuan Zhang
I am a Ph.D. student in Data Science at UC San Diego, advised by Prof. Hao Zhang. My research focuses on systems for LLM inference and reinforcement learning, including efficient model execution and infrastructure for training agents at scale.
Previously, I interned at Google on the Tunix team. I received my B.S. in Data Science from UC San Diego, where I worked with Prof. Hao Zhang and Prof. Haojian Jin on language-model agents and reinforcement learning.
Email: yuz165 [at] ucsd [dot] edu / GitHub / X / LinkedIn / Google Scholar
Recent News
Selected Work
FastAFD — Attention–FFN Disaggregation for MoE Inference
FastAFD separates KV-cache-constrained attention from MoE expert execution, aggregating tokens across attention workers to keep expert batches large. Its runtime overlaps communication, computation, and scheduling to make disaggregation effective on GB200 NVL72. On Qwen3-235B-A22B-FP8 and MiniMax-M2.5, it achieves 1.35–1.45× per-GPU steady-state decode throughput over a tuned colocated vLLM baseline.
GamingAgent & GRL — Game-Based Evaluation and Multi-Turn RL
GamingAgent combines standardized game environments with a modular perception–memory–reasoning harness, enabling comparisons between bare LLM/VLM policies and agentic workflows. GRL extends this ecosystem from evaluation to multi-turn reinforcement learning, with training backends built on VERL and Tunix.
NanoRollout — Scaling Digital Agent RL and Distillation
We use NanoRollout to study how rollout scale and training-data composition affect digital agents. In matched-recipe SWE-agent experiments, scaling rollout batches to 4,096 improves RL performance. Our distillation study also exposes a key failure mode: single-harness training can perform worse than the base model under an unseen harness. These findings motivate multi-harness trajectory mixtures, used to train Mocha-Coder-32B.
Tunix — LLM Post-Training on TPUs
During my Google internship, I implemented sequence packing for LLM training on TPUs, reducing padding overhead and improving training efficiency.