MRCL

Researchers introduced the MRCL benchmark to test reinforcement learning's effectiveness in continual post-training for vision-language models, finding that RL alone does not prevent catastrophic forg…