OR Else: A Differentiable Trust Region for Policy Optimization

En 30 secondes

Ce qui s'est passé

Researchers propose Output Reset, a smooth alternative to clipped objectives in PPO and GRPO for language model training, replacing abrupt gradient changes with squared-margin loss.

Pourquoi ça compte

Matters for engineers optimizing reinforcement learning training stability in large language models, particularly when fine-tuning with policy gradient methods.

Vigilance

Results are mixed: PPO-OR shows gains under GAE but GRPO-OR does not improve reward scores at group size two. Scores measure training-time reward models, not held-out human preferences.