From the Inside Out: Progressive Distribution Refinement for Confidence Calibration
This work addresses reward hacking and calibration issues in RL, which is important for improving model reliability, but appears incremental as it builds on existing test-time scaling strategies.
The paper tackles the problem of reward hacking and test-training discrepancy in reinforcement learning by proposing DistriTTRL, which uses distribution priors and diversity penalties to optimize self-reward signals, achieving significant performance improvements across multiple models and benchmarks.
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.