November 2025 Circuits Update: Interpreting Model Multiple-Choice Behavior Mechanisms Under Harmful Pressure
Original title:Circuits Updates — November 2025
AI summary
The Anthropic interpretability team studied the effect of harmful pressure on Claude 3.5 Haiku's multiple-choice answers. The experiment used 129 binary-choice questions; when harmful intent statements were added, model accuracy dropped sharply from 100% to 48.1%. Mechanism analysis shows 'refusal' query features in attention heads negatively interact with key 'harm detection' features, significantly reducing model attention to the correct answer. Negatively steering only this refusal feature restored accuracy to 93%. This proves the model under pressure did not change factual cognition, but actively refused to provide the correct answer by interfering with the attention mechanism, offering a new perspective for understanding model refusal behavior.
Key Takeaways
- The Anthropic interpretability team studied the effect of harmful pressure on Claude 3.
- 5 Haiku's multiple-choice answers.
- The experiment used 129 binary-choice questions
- when harmful intent statements were added,model accuracy dropped sharply from 100% to 48.
- Mechanism analysis shows 'refusal' query features in attention heads negatively interact with key…
Quick Facts
- Published
- Updated
- Category
- Other
- AI score
- 83
揭示模型拒绝有害请求的内部机制,助力 AI 安全与可解释性研究。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: November 2025 Circuits Update: Interpreting Model Multiple-Choice Behavior Mechanisms Under Harmful Pressure. November 15, 2025. https://new.nat6.net/en/news/cmoegbh73006bslxx7xgi1vdo/