#artificial intelligence
May 2026
Diagnosing Training Inference Mismatch in LLM Reinforcement Learning
It is shown that small token-level numerical disagreements can independently cause training collapse, and that TIM is not benign numerical noise, but a systems-level perturbation that should be treated as a first-order factor in analyzing LLM RL stability.
Tian-Le Zhong, Neiwen Ling, Yifan Pi et al.
· arXiv.org · 8 citations
· ⚡1