A single vector in the mid-to-late residual stream of real Llama3-8b-Instruct is differentially activated during correct self-written-text recognition, is causally related to the model's ability to perceive and assert self-authorship; its decision-relevant role is not found in the corresponding base model

- Ackerman & Panickssery find that Llama3-8b-Instruct -- but not the base Llama3-8b model -- reliably distinguishes its own outputs from human-written text, and identify a vector in the residual stream that is differentially activated when the model makes a correct self-written-text recognition judgment. [ackerman-panickssery-2024-self-generated-text-recognition] - The vector is causally related to the model's ability to both perceive and assert self-authorship: applying it to the model's own generated output as it is produced steers the model to claim or disclaim authorship at will, and applying it to arbitrary texts as the model reads them steers the model to believe or disbelieve it wrote them. [ackerman-panickssery-2024-self-generated-text-recognition] - The vector's effect is fairly specific to self-authorship (positive steering has no effect on an unrelated name-recognition "dummy" task, though negative steering with a high multiplier has a modest effect; zeroing it out has no effect on the dummy tasks), and for the base Llama3-8b model the vector's activation does not track the model's authorship probabilities and projecting it out has no effect. The vector (derived from the chat model) is nonetheless more activated by self-written than other-written texts in the base model too, so the authors' interpretation is that the base model carries the self-text information but only the chat model converts it into a decision, consistent with the vector being related to post-training exposure to its own outputs. [ackerman-panickssery-2024-self-generated-text-recognition]