让下一篇重要论文，不再错过。

原文摘要

Reward models (RMs) trained from human preferences are central to aligning large language models, yet they often break under distribution shift or targeted perturbations. Existing failure discovery methods rely on prior knowledge of preference attributes and therefore do not scale to new models or data. We introduce a preference distribution agnostic procedure that uses the reward model itself to guide controlled decoding toward mis specified responses while preserving the underlying preference class. Building on this discovery mechanism, we propose REFORM, a self improving RM framework that (i) searches for class consistent but reward inconsistent variants and (ii) fine tunes the RM on a small, targeted augmentation of these failures. On Anthropic Helpful Harmless and PKU Beavertails, REFORM consistently improves robustness without degrading in distribution reward quality across different models (e.g., Mistral-7B and Qwen-14B), with an average improvement of 35%–45%.Further, across Best of N sampling, PPO, and DPO, REFORM preserves downstream generation quality and reduces spurious correlations. Our results show that RMs can serve as their own adversary to expose and fix blind spots, yielding robust alignment without manual attribute priors or large scale relabeling.