Reproduce “Diagnosing Training Inference Mismatch in LLM Reinforcement Learning via a Zero-Mismatch Reference” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianle Zhong, Neiwen Ling, Yifan Pi, Zijun Wei, Tianshu Yu, Geoffrey Fox, Peng Wu, Xiao Yu
VeXact provides a reference for studying training–inference mismatch (TIM) in LLM reinforcement learning, aligning its model and kernels with the trainer through deterministic, batch-invariant kernels. In controlled REINFORCE experiments on DAPO, Qwen3-30B-A3B’s vLLM training reward fell from 0.574 to 0.255, while VeXact reached 0.753. These math-focused results do not establish generality beyond tested tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.