A mechanistic study of language model introspection.

arXiv:2609.35108v1 Announce Type: cross Abstract: Large language models (LLMs) can sometimes report perturbations to their internal activations---even when the input provides no evidence that an intervention occurred. How do models detect and localize such int