From the Inside Out: Progressive Distribution Refinement for Confidence Calibration

TL;DR

This paper introduces DistriTTRL, a novel reinforcement learning method that uses distribution priors of model confidence for progressive reward optimization and employs diversity penalties to prevent reward hacking, leading to improved performance.

Contribution

DistriTTRL is the first approach to leverage confidence distribution priors in RL and mitigate reward hacking through diversity penalties, enhancing training stability and effectiveness.

Findings

Significant performance improvements across multiple benchmarks.

Effective mitigation of reward hacking issues.

Enhanced model training through confidence distribution refinement.

Abstract

Leveraging the model's internal information as the self-reward signal in Reinforcement Learning (RL) has received extensive attention due to its label-free nature. While prior works have made significant progress in applying the Test-Time Scaling (TTS) strategies to RL, the discrepancy in internal information between test and training remains inadequately addressed. Moreover, Test-Time Training based on voting-based TTS strategies often suffers from reward hacking problems. To address these issues, we propose DistriTTRL, which leverages the distribution prior of the model's confidence during RL to progressively optimize the reward signal, rather than relying solely on single-query rollouts. Additionally, we mitigate the phenomenon of consistent reward hacking caused by the voting-based TTS strategies through diversity-targeted penalties. Benefiting from this training mechanism where model capability and self-reward signals complement each other, and the mitigation of reward hacking, DistriTTRL has achieved significant performance improvements across multiple models and benchmarks.