Learning, Fast and Slow: LLMs That Adapt Without Forgetting

Catastrophic forgetting and plasticity loss are two separate costs of RL post-training. This paper interleaves slow weight updates (RLVR / CISPO) with a fast, prompt-evolving channel (GEPA) — reaching RL's peak accuracy with up to 3× fewer samples, drifting up to 70% less from the base model, and staying plastic enough to learn the next task where pure RL stalls.