Persona Without Substrate: Regime-Dependence and the LLM Individuation Problem

Short summary

ArXiv paper challenges the assumption that LLM persona directions remain consistent across prompting, fine-tuning, and inference-time steering contexts. Experiments on Qwen and Mistral show the same vector yields different semantic content depending on regime. Authors propose regime-indexed individuation: identity should bind direction AND context, not direction alone.

• •Persona-vector literature assumes same direction produces same content across contexts
• •Empirical evidence shows contradictions in prompt-extracted versus fine-tuned versus inference-steered personas
• •Proposes identity should be (vector, regime) pairs for consistent regime-internal behavior

Generated with AI, which can make mistakes.