Reinforcement learning with verifiable rewards (RLVR) is a common approach for improving reasoning in language models. A new arXiv preprint, 'Where RL Breaks: Localizing and Repairing Failures via Targeted Continued Pretraining,' examines a known limitation: RLVR often sharpens skills the base model already has rather than adding new ones.
The abstract states that growing evidence points to this sharpening effect, and the paper's title indicates a proposed solution. The authors appear to focus on localizing specific failures of RL and then applying targeted continued pretraining to repair those failures, rather than relying on broader reinforcement learning alone.
Because the abstract is truncated, the full method and experimental results are not available from the source. The preprint is the only source for this digest, so the summary is limited to the stated premise and the direction implied by the title.