Title: The Assistant as a Canonical Reference in Cross-Persona Self-Recognition

Abstract

Post-trained language models possess the ability to identify their own outputs, even when presented with only a sentence or two devoid of context. In a companion study \citep{jack2026twomodes}, we demonstrated that these models can also detect when they are operating "on-policy," evidenced by a distinct drop in entropy during assistant-mode generation. Both of these phenomena are linked to the Assistant persona, which is primarily defined during post-training. This study expands the scope to examine cross-persona authorship judgments using Llama-3.1-70B-Instruct. We construct a matrix tracking authorship claim rates across a diverse set of evaluator and generator personas, ranging from librarians to dragons and Shakespeare. Our analysis yields two primary findings.

First, we observe a tight coupling on the Assistant’s specific row within the matrix: the Assistant’s claim rate correlates strongly with the persona-vector distance from the Assistant in activation space, as well as the entropy gap between the Assistant’s surprise at a given persona’s text and that persona’s surprise at its own text. This finding generalizes the entropy signature of acting (reported in the companion paper) into a retrospective signature of having acted.

Second, this coupling breaks down outside the Assistant’s row. The natural symmetric extension of the entropy gap fails to predict authorship for distinctive evaluators such as pirates, dragons, or Shakespeare. Instead, prediction relies on an asymmetric metric: the evaluator’s surprise relative to the Assistant’s surprise on the same text, rather than the generator’s surprise. We tested numerous candidate substitutes to rule out the possibility that any persona could serve as this reference point, but none succeeded. We interpret this asymmetry as the model conducting an implicit Bayesian likelihood-ratio test against the Assistant as the canonical alternative hypothesis. The persona-vector geometry established by \citet{chen2025persona}, where every persona is represented as a delta from the Assistant, ensures that the Assistant remains the only persona universally accessible to this test.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC