Xingjian Zhang 张行健

Meta Superintelligence Labs

Muse Product Post-Training

• Co-developed steering mechanisms for agent rollouts, including reflection, plan-and-execute, and adaptive replanning; migrated experimentation to a controlled white-box sandbox for reproducible iteration.
• Contributed personalized-research SFT data to Muse Spark post-training; the trained model improved by 3pp across three internal agent benchmarks.

Google DeepMind

Gemini Post-Training

• Curated a benchmark linking YouTube Shorts to real-world events to evaluate video grounding through web search.
• Contributed agentic RL and multimodal-input support to simply ↗, a minimal JAX codebase used broadly within DeepMind with 500+ GitHub stars.

Google DeepMind

Gemini Post-Training

• Proposed GPU/TPU-friendly adaptive Transformer layer skipping based on input-sequence complexity.
• Matched full-compute performance on language and sequential user modeling at 60% compute; integrated the method into TensorFlow Recommenders. paper ↗

Ph.D. in Information Science

University of Michigan, School of Information · advised by Prof. Qiaozhu Mei

B.S. in Computer Science; Minor in Mathematics

University of Michigan, EECS Department