RL Forgets! Towards Continual Policy Optimization

Name: RL Forgets! Towards Continual Policy Optimization
Uploaded: 2026-10-04T00:06:24.785118+00:00
Duration: 325 s
Description: Standard reinforcement learning methods fail to prevent catastrophic forgetting when vision-language models are continually post-trained. This research introduces MRCL, a new benchmark, and Continual Policy Optimization (CPO), a replay-free framework. CPO solves the problem by replacing an impossible constraint on old data with a parameter movement regularization technique to limit policy drift.

Standard reinforcement learning methods fail to prevent catastrophic forgetting when vision-language models are continually post-trained. This research introduces MRCL, a new benchmark, and Continual Policy Optimization (CPO), a replay-free framework. CPO solves the problem by replacing an impossible constraint on old data with a parameter movement regularization technique to limit policy drift.

Open the full episode