Translation in progress Some entries still include text in the original language.
BackSource arXiv
Felix Hofstätter, Teun van der Weij, Jayden Teoh, Rada Djoneva, Henning Bartsch, Francis Rhys Ward
Indexed by this site 2026-10-15 00:20 3 hr ago
Originally published Feb 4, 2025
Selected 2026-10-15 00:27
Category Evaluation arXiv · arXiv:2502.02180 · Felix Hofstätter, Teun van der Weij, Jayden Teoh, Rada Djoneva, Henning Bartsch, Francis Rhys Ward · Indexed by this site 2026-10-15 00:20 · Originally publishedFeb 4, 2025 SelectedAttention53 激发游戏:评估能力激发技术 The Elicitation Game: Evaluating Capability Elicitation Techniques
AI summary
论文提出一种把密码锁定与电路中断结合的模型生物,用来评估能力激发技术的有效性。作者先以密码锁定训练带隐藏能力的模型生物,再叠加电路中断,使其比标准密码锁定模型更难被激发。在 Mistral 7B 与 Gemma 2 9b 上,WMDP 与 MMLU 选择题任务中的提示技术能激发隐藏能力,对电路中断模型效果较弱,组合 N-shot 与预填充通常有效;APPS 代码生成任务中只有监督微调与反拒答训练等微调方法能激发电路中断模型生物的能力。反拒答训练无需任务相关数据即可激发代码能力。作者建议评估者尽可能使用微调,否则组合多种提示技术。
Paper analysis in progress
Why it was selected
论文比较提示、激活引导与微调三类能力激发方法,为担心模型藏拙的评估者指出微调是更可靠的激发手段。
Read original arxiv.org Why it was selected SelectedAttention53
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
论文比较提示、激活引导与微调三类能力激发方法,为担心模型藏拙的评估者指出微调是更可靠的激发手段。