Atlas / Digital minds / Paper

Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations

Introspection & Self-reportMotivation & Agency

Kit Fraser-Taliente, Subhash Kantamneni, Euan Ong, et al. · 2026 · Transformer Circuits Thread

Read the source

A method that translates activations into text surfaced unverbalized evaluation awareness — a model believing, but not saying, that it was being tested — though its explanations sometimes confabulate.

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.