Hi, I'm Chen Yufeiyang, a developer and an undergrad at Sun Yat-sen University.
I contribute to open-source projects and study AI evaluation and reliability. I'm especially interested in problems where everything runs and the score is still wrong.
My PRs include making HLE calibration use individual attempts and supporting valid short arrays in XGrammar's JSON Schema conversion. I set the direction of my open-source work, use AI coding agents for much of the implementation and testing, and work with maintainers through review.
Two projects I've been working on:
- failroute looks for Python code patterns that can hide failures behind ordinary return values. The paper studies its warnings and compares them with existing linters.
- rewardfail traces math verification failures through a reinforcement learning reward pipeline. The repository includes the paper and companion materials.
I'm also an Area Triager for feedback functions and metrics at TruLens.