Score centering reduces instability in LLM RL training

Recent research explores score centering as a practical method to improve training stability in large-scale reinforcement learning for large language models. The approach addresses training-inference mismatch caused by discrepancies between sampler and trainer policies in disaggregated architectures. Unlike high-variance importance sampling techniques, score centering provides an additive correction to eliminate bias without requiring destructive clipping or truncation. Open article →