Study examines training-inference mismatch in LLM reinforcement learning
A new arXiv paper analyzes how mismatches between training and inference arise in on-policy reinforcement learning for large language models, where training assumes responses come from the current policy but inference engines differ in practice.
Read the full story at arXiv cs.LG ↗
Timeline · 1 report
- 2026-11-02 05:00 · arXiv cs.LG
On the Origin and Evolution of Training-Inference Mismatch in LLM Reinforcement Learning