CAMEL: Confidence-Gated Reflection for Reward Modeling
This addresses the efficiency-interpretability trade-off in reward modeling for aligning LLMs with human preferences, representing a strong incremental improvement.
The paper tackles the trade-off between efficiency and interpretability in reward models for aligning large language models with human preferences by proposing CAMEL, a confidence-gated reflection framework that uses log-probability margins to gauge instance difficulty and selectively invokes reflection. It achieves state-of-the-art performance with 82.9% average accuracy on three benchmarks, surpassing prior models by 3.2% and outperforming larger models using only 14B parameters.
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.