CAMEL: Confidence-Gated Reflection for Reward Modeling

TL;DR

CAMEL introduces a confidence-gated reflection framework for reward modeling that improves accuracy and efficiency by selectively invoking reflection based on instance difficulty, using reinforcement learning for self-correction.

Contribution

It proposes a novel confidence-based decision mechanism and reinforcement learning training method to enhance reward model performance and interpretability.

Findings

Achieves 82.9% average accuracy on reward-model benchmarks.

Surpasses prior models by 3.2% in accuracy.

Outperforms larger models with fewer parameters, establishing a better accuracy-efficiency trade-off.

Abstract

Reward models play a fundamental role in aligning large language models with human preferences. Existing methods predominantly follow two paradigms: scalar discriminative preference models, which are efficient but lack interpretability, and generative judging models, which offer richer reasoning at the cost of higher computational overhead. We observe that the log-probability margin between verdict tokens strongly correlates with prediction correctness, providing a reliable proxy for instance difficulty without additional inference cost. Building on this insight, we propose CAMEL, a confidence-gated reflection framework that performs a lightweight single-token preference decision first and selectively invokes reflection only for low-confidence instances. To induce effective self-correction, we train the model via reinforcement learning with counterfactual prefix augmentation, which exposes the model to diverse initial verdicts and encourages genuine revision. Empirically, CAMEL achieves state-of-the-art performance on three widely used reward-model benchmarks with 82.9% average accuracy, surpassing the best prior model by 3.2% and outperforming 70B-parameter models using only 14B parameters, while establishing a strictly better accuracy-efficiency Pareto frontier.