TRL Adds Score Centering to Stabilize RL Training

Hugging Face's TRL library introduced Score Centering to address instability in LLM RL training caused by mismatched training and sampling policies, and has now made it the default along with masked importance sampling in the async GRPO Trainer.

2026-10-27 ~ 2026-10-27 · 2 related posts

• Hugging Face TRL makes score centering + masked importance sampling default in async GRPO — _lewtun · 2026-10-27
• TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27