LLMs reconstruct the current persona at least in part via attention to past persona activations stored in the KV cache

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

• We hypothesize that axes of persona differentiation within LLMs are likely already present in base models and inherited from the pre-training corpushypothesis0.780
  Motivated by near-identical PCs for base and instruct Gemma
• The leading component of the persona space of instruct LLMs is an 'Assistant Axis' that captures the extent to which a model is operating in its default Assistant modeclaim0.773
  Primary empirical claim of the paper
• Are mind-like states and mechanisms in LLMs that operate in persona-relative ways controlled by persona vectors?question0.772
  Identified as an open research direction for future mechanistic interpretability work
• This study is the first to explore subtle OOC behavior in persona-assigned LLMs during open-ended text generationclaim0.770
  Novelty claim establishing the paper's contribution relative to prior work focused on closed-form tasks
• LLMs demonstrate stronger persona fidelity for clearly defined and socially desirable high-level personas than for neutral or low-level personasclaim0.769
  Observed across multiple models and tasks; attributed to RLHF training preference for helpful/harmless/honest responses
• Persona Assignment to LLMsconcept0.767
  The practice of providing LLMs with a persona description to shape their generated responses
• We suspect it has to do with personas being latent factors that persist for many tokens; thus, recent expression of a persona should predict its near-future expression.quote0.763
  Load-bearing mechanistic conjecture about why persona vectors generalize from extraction to prediction
• Persona representations originate in pretraining, not alignment; they form without explicit supervision as useful features for next-token predictionclaim0.758
  Core interpretive claim providing mechanistic explanation for early persona formation