Handling Hard Boundaries in Off-Policy Training

ziv_ravid · x · 2026-07-11

This segment covers an off-policy training approach: they discard the old policy and use the logprobs of the "current policy / rollout engine" as the importance ratio. Any samples exceeding the two-sided hard boundaries are directly masked out in the gradient, rather than applying standard clipping.

The core idea is that in agent training, they use stricter filtering to control off-policy errors instead of relying on mild clipping.

Related event: GLM Team Proposes SAO Algorithm for Asynchronous Agent RL(15 posts)→

Original post →

More from Research

Research channel →