Jian Yang studies how language models write and evaluate code. He co-authored Qwen2.5-Coder at Alibaba and leads LoopCoder research at Beihang University.[1][4][3]
Start with LoopCoder-v2 →
GitHub · CSJianYang Career
Contributions
LoopCoder-v2 passes a model’s internal representation through the same layers again before it emits a token. Yang and his co-authors compare models trained with different loop counts to measure what each extra pass adds and where the gains taper off.[3]
Tools, writing & talks
Writing, research & talks
Start here
LoopCoder-v2arxiv.orgA study of how repeated passes through shared layers affect coding performance and inference cost.
Start with the comparison across loop counts to see where extra inference computation pays off.
First listed author[3]- Qwen2.5-Coder Technical Reportarxiv.org
The training and evaluation report for Alibaba’s open code-model family.
Co-author[4]