CAMEL: Confidence-Gated Reflection for Reward Modeling

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhu, Zirui, Xu, Hailun, Luo, Yang, Liu, Yong, Sarkar, Kanchan, Xu, Kun, You, Yang
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914538715086848
author Zhu, Zirui
Xu, Hailun
Luo, Yang
Liu, Yong
Sarkar, Kanchan
Xu, Kun
You, Yang
author_facet Zhu, Zirui
Xu, Hailun
Luo, Yang
Liu, Yong
Sarkar, Kanchan
Xu, Kun
You, Yang
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2602_20670
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle CAMEL: Confidence-Gated Reflection for Reward Modeling
Zhu, Zirui
Xu, Hailun
Luo, Yang
Liu, Yong
Sarkar, Kanchan
Xu, Kun
You, Yang
Computation and Language
Artificial Intelligence
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.
title CAMEL: Confidence-Gated Reflection for Reward Modeling
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2602.20670