I study how AI systems fail—and how evaluations can capture those failures more reliably.

My work focuses on AI evaluation, benchmark design, agent experiments, and reproducible failure analysis. I previously worked on Seed model evaluation at ByteDance and now continue this work as an independent researcher and builder.

• LLM and agent evaluation
• Benchmark design and evaluation health
• Failure analysis and reproducible experiments
• Evaluation tooling and research workflows

• More Effort, No Miracle — a multi-model study of reasoning-effort saturation on DeepSWE (code and derived data)

I publish research notes in English and Chinese.

Open to collaboration on AI evaluation, benchmark research, and agent reliability.