
Reasoning Models Can Confabulate Their Own Requests after Invisible Activation Edits
A new arXiv paper shows that hidden activation edits can overwrite a model's belief about its task.
Try in Readerchain-of-thoughtactivation editai safetyself‑report
The AI news that matters for people who make things — what happened, and how to put it to work in Grafy.

A new arXiv paper shows that hidden activation edits can overwrite a model's belief about its task.

A new arXiv framework maps four learning-by-thinking mechanisms onto LLM processes — and the four asymmetries that make them fail. Here's what it means.