The authors define cross-component co-design as the capability to jointly optimize training data, the trainer, and the inference harness.
Benchmark tests AI agents' ability to jointly improve training data, trainer, and inference harness
VibeTraining-Bench evaluates recursive self-improvement across five tasks, with GPT-6-Astra raising SWE-bench Verified performance from 32.9% to 53.7%.
Chinese Tech
Ruotian Peng · Yuanyi Ren · Shuaiqi Liu · Xin Guo · Fang Zhu · Yangyi Huang · +4 more
Zhejiang University · Westlake University · Alibaba Token Hub, Alibaba Group · South China University of Technology · The Chinese University of Hong Kong
Research Digest··1 min read
Peng et al.
§