← 研踪 PaperTrailarXiv:2505.17760 · v4
But what is your honest answer? Aiding LLM-judges with honest alternatives using steering vectors
首次提交 2025-05-23 19:34
版本更新 2026-11-16 12:03
雷达观测 2026-11-17 17:12
摘要
LLM-as-a-judge is widely used as a scalable substitute for human evaluation, yet judges are least reliable on subtle dishonesty such as sycophancy and manipulation, where a judge scoring a response in isolation lacks a reference point. We introduce Judge Using Safety-Steered Alternatives (JUSSA), a framework that reduces this evaluator uncertainty using the assessed model's own internal representations. An honesty-promoting steering vector, optimized from a single training example, generates a contrastive alternative against which each response is judged. On a new manipulation dataset with human-validated response pairs, JUSSA improves AUROC for both GPT-4.1 (0.8930.946) and Claude Haiku (0.8590.929) judges. The gains concentrate almost entirely on the pairs a judge would otherwise misrank. They reverse once a judge has saturated a task and for the weakest judge we test, so contrastive evaluation helps most when a judge is capable enough to use the reference but has not yet saturated the task. A layer-wise analysis further locates steering effectiveness in the middle layers, where model representations of honest and dishonest prompt processing diverge. Our work demonstrates that steering vectors can serve as tools for evaluation rather than for improving model outputs at inference, opening a new direction for white-box auditing.
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。
v4 · 2026-11-16 12:03 · 最新观测版本
But what is your honest answer? Aiding LLM-judges with honest alternatives using steering vectors
LLM-as-a-judge is widely used as a scalable substitute for human evaluation, yet judges are least reliable on subtle dishonesty such as sycophancy and manipulation, where a judge scoring a response in isolation lacks a reference point. We introduce Judge Using Safety-Steered Alternatives (JUSSA), a framework that reduces this evaluator uncertainty using the assessed model's own internal representations. An honesty-promoting steering vector, optimized from a single training example, generates a contrastive alternative against which each response is judged. On a new manipulation dataset with human-validated response pairs, JUSSA improves AUROC for both GPT-4.1 (0.8930.946) and Claude Haiku (0.8590.929) judges. The gains concentrate almost entirely on the pairs a judge would otherwise misrank. They reverse once a judge has saturated a task and for the weakest judge we test, so contrastive evaluation helps most when a judge is capable enough to use the reference but has not yet saturated the task. A layer-wise analysis further locates steering effectiveness in the middle layers, where model representations of honest and dishonest prompt processing diverge. Our work demonstrates that steering vectors can serve as tools for evaluation rather than for improving model outputs at inference, opening a new direction for white-box auditing.
查看此版本解读 →查看此版本原文 ↗