Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards

TL;DR

This paper introduces DCPO, a framework that decouples reasoning and calibration in reinforcement learning for large language models, improving calibration without sacrificing accuracy.

Contribution

It reveals a gradient conflict in existing methods and proposes a decoupling approach that enhances calibration while maintaining accuracy.

Findings

DCPO achieves superior calibration performance.

DCPO preserves accuracy comparable to existing methods.

It substantially reduces over-confidence in LLMs.

Abstract

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.