Search papers, labs, and topics across Lattice.
This paper identifies a gradient conflict between optimizing for policy accuracy and calibration in Reinforcement Learning from Verifiable Rewards (RLVR). To address this, they propose Decoupled Policy Optimization (DCPO), a framework that separates the reasoning and calibration objectives during training. Experiments show that DCPO maintains accuracy comparable to GRPO while significantly improving calibration and reducing overconfidence in LLMs.
LLMs trained with reinforcement learning become overconfident in wrong answers due to a fundamental conflict between accuracy and calibration objectives, but this can be fixed by decoupling these objectives during training.
Reinforcement Learning from Verifiable Rewards (RLVR) significantly enhances large language models (LLMs) reasoning but severely suffers from calibration degeneration, where models become excessively over-confident in incorrect answers. Previous studies devote to directly incorporating calibration objective into existing optimization target. However, our theoretical analysis demonstrates that there exists a fundamental gradient conflict between the optimization for maximizing policy accuracy and minimizing calibration error. Building on this insight, we propose DCPO, a simple yet effective framework that systematically decouples reasoning and calibration objectives. Extensive experiments demonstrate that our DCPO not only preserves accuracy on par with GRPO but also achieves the best calibration performance and substantially mitigates the over-confidence issue. Our study provides valuable insights and practical solution for more reliable LLM deployment.
