When Activation Oracles Learn Not to Read: Concept-Specific Blind Spots in Fine-Tuned Oracles

• 类型：arxiv
• 标识：2607.23379
• 链接：https://arxiv.org/abs/2607.23379
• 主分类：engineering
• 形态：method
• 被引：1
• 被引来源：Semantic Scholar
• S2被引：1
• 影响力被引：0
• TLDR：It is found that fine-tuned AOs can become concept-specific anti-readers: they selectively fail to recover the concept persistently present during their own training, raising a reliability concern for learned interpretability interfaces.
• 待LLM分类：否
• 标题中文：当激活预言机学会不去读取：微调预言机中的概念特定盲区
• TLDR中文：研究发现，微调后的 AO 可能变成概念特异的 anti-reader：它们会选择性地无法恢复自身训练过程中持续存在的概念，从而对习得的可解释性接口提出可靠性担忧。
• 来源文件：
• /inbox/tom/_candidates/2026-08-10-agent-rag-longcontext-candidates.json
• [S2 enrich]