Atlas / Digital minds / Paper

Mechanisms of Introspective Awareness

Introspection & Self-report

Uzay Macar, Li Yang, Atticus Wang, Peter Wallich, Emmanuel Ameisen, Jack Lindsey · 2026 · arXiv preprint

Read the source

Detecting an injected concept runs through a two-stage circuit: early features carrying the evidence suppress later features that would otherwise answer ‘no’. The circuit appears only after preference-based post-training, and removing refusal directions raised detection without false positives.

Questions it bears on

Help improve this map

Know a paper that belongs here, or think something needs correcting?

Suggest a correction ↗Contributions are reviewed before they appear.