WeWorkBuddy

But what is your honest answer? Aiding LLM-judges with honest alternatives using steering vectors

arXiv:2505.17760v4 Announce Type: replace-cross Abstract: 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 vecto

阅读原始来源
来源参考:But what is your honest answer? Aiding LLM-judges with honest alternatives using steering vectors。暂未取得截图。
来源记录 · 预览待采集 · arxiv.org打开来源查看完整网页。

来源记录

项目分析

暂无五维分析,可先查看上方来源资料。