Other

Emergent Introspection in Large Language Models

Original title:Emergent Introspective Awareness in Large Language Models

AI summary

Research uses a 'concept injection' technique to directly manipulate model internal activation states to test whether LLMs have true introspection. Experiments find that in some contexts, the model can detect and identify injected concepts, distinguish its own internal representations from original text input, and even use recall of prior intentions to distinguish its own outputs from artificially preset content. Among them, the Claude Opus series shows the strongest introspection, but this ability is unstable and highly context-dependent. Research shows current models have some functional awareness of their internal states; although still unreliable, it may develop as model capabilities improve.

Key Takeaways

  • Research uses a 'concept injection' technique to directly manipulate model internal activation…
  • Experiments find that in some contexts,the model can detect and identify injected…
  • Among them,the Claude Opus series shows the strongest introspection,but this ability is unstable…
  • Research shows current models have some functional awareness of their internal states
  • although still unreliable,it may develop as model capabilities improve.

Quick Facts

Published
Updated
Source
Anthropic:Transformer Circuits(可解释性研究)
Category
Other
AI score
83
Read the original Read the full article at the original source
Why it matters

研究揭示大模型可能具备有限内省能力,对 AI 安全和透明度有重要启示。

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.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: Emergent Introspection in Large Language Models. October 29, 2025. https://new.nat6.net/en/news/cmoegbh73006cslxx667v4iig/

View all →