What happens to the persona during user turns, when the model is processing input rather than generating?

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• The persona region is not continuously maintained during user turns but is active only while the model is producing its own tokensclaim0.813
  Finding from mini experiment 1 that the assistant axis tracks the user independently of the current generation persona during user tokens
• The model-persona view crucially depends on Hypothesis 3; without discrete persona regions, the simulator view is more aptclaim0.779
  Identifies the key theoretical vulnerability of the model-persona view
• Ududec et al. 2026 find that once a model has entered the evil persona region over the course of a conversation, it is difficult to steer it backfinding0.779
  Evidence that the evil persona region exhibits the stickiness hallmark of a genuine attractor basin
• Mini experiment 1: During user turns, assistant-capped and uncapped activation traces along the assistant axis are nearly identical in Qwen 3 32B, indicating the persona is not continuously maintained during user token processingfinding0.779
  Preliminary finding from the authors' own experiment supporting claim about persona gap during user turns
• A single direction encoding whether the model represents a claim as true or false is persona-relative across conversations (Lampinen et al. 2026)finding0.778
  Additional evidence that core representations are persona-relative, supporting Claim about persona-relative representations
• How does different post-training data shift a model's position along persona dimensions?question0.776
  Future work direction: using persona space to study effects of training data on model character
• How does the persona persist through token-time across both generation and user turns?question0.776
  Motivating question for mini experiment 2 about the mechanism of persona continuity
• Persona representations originate in pretraining, not alignment; they form without explicit supervision as useful features for next-token predictionclaim0.776
  Core interpretive claim providing mechanistic explanation for early persona formation