Other

Circuits Update — October 2025

Original title:Circuits Updates — October 2025

AI summary

The Anthropic interpretability team shared multiple research advances. Research found cross-modal visual features in models from Haiku 3.5 to Sonnet 4.5, able to recognize semantic concepts encoded in ASCII art and SVG code, such as eyes, mouths, dogs, cats, etc. These features depend on the context of visual depiction; for example, an SVG circle element activates the 'eye' feature only when located in a larger structure activating 'face' features. Steering some features during generation can correspondingly modify the semantics of text art, e.g., turn an ASCII emoticon from frown to smile, or add wrinkles to an SVG face. Research also found a model tendency similar to 'face pareidolia', interpreting shapes as components of animal drawings. These features also work for human-drawn SVG.

Key Takeaways

  • The Anthropic interpretability team shared multiple research advances.
  • Research found cross-modal visual features in models from Haiku 3.
  • 5 to Sonnet 4.
  • 5,able to recognize semantic concepts encoded in ASCII art and SVG code,such as…
  • These features depend on the context of visual depiction

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

为AI可解释性研究提供新实验方法,启发跨模态模型设计。

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: Circuits Update — October 2025. October 15, 2025. https://new.nat6.net/en/news/cmoegbh73006dslxxjjzsxq3s/

View all →