← 研踪 PaperTrailarXiv:2611.20955 · v1
Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning
首次提交 2026-11-16 20:01
版本更新 2026-11-16 20:01
雷达观测 2026-11-17 17:11
摘要
Probabilistic safety judges return the probability of a safety label and cheaply score reinforcement learning rollouts. A confidence gate sends only uncertain verdicts to a stronger LLM judge, which leaves confidently passed harm in a blind spot. Policies can exploit this blind spot, and commercial judges release new versions. We propose a protocol that validates such a judge before and during training, separates the share of confidently passed harm from its amount, and re-validates it after each judge change. We apply it to two versions of Jev, a commercial decision-only judge, and to an open guard classifier. Under the probability reward of Jev, passed harm shrinks in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrates in the blind spot. Binarizing this judge meets the escape criterion on Llama, and the open guard classifier shows an escape signal in one seed under a second reference. Guard classifiers share this blind spot, and even frontier LLM judges catch at most one of 10 human-labeled confident misses. Without a trained reward model, the probability reward with a helpfulness term lowers HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.
作者备注:15 pages, 2 figures
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-16 20:01 · 最新观测版本
Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning
Probabilistic safety judges return the probability of a safety label and cheaply score reinforcement learning rollouts. A confidence gate sends only uncertain verdicts to a stronger LLM judge, which leaves confidently passed harm in a blind spot. Policies can exploit this blind spot, and commercial judges release new versions. We propose a protocol that validates such a judge before and during training, separates the share of confidently passed harm from its amount, and re-validates it after each judge change. We apply it to two versions of Jev, a commercial decision-only judge, and to an open guard classifier. Under the probability reward of Jev, passed harm shrinks in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrates in the blind spot. Binarizing this judge meets the escape criterion on Llama, and the open guard classifier shows an escape signal in one seed under a second reference. Guard classifiers share this blind spot, and even frontier LLM judges catch at most one of 10 human-labeled confident misses. Without a trained reward model, the probability reward with a helpfulness term lowers HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.
查看此版本解读 →查看此版本原文 ↗