Preprint argues persona representations in LLMs are regime-dependent, not cross-regime invariant

A new preprint on arXiv argues that persona representations in large language models (LLMs) are not invariant across different operational regimes, challenging a foundational assumption in the persona-vectors literature.

The paper, titled “Persona Without Substrate: Regime-Dependence and the LLM Individuation Problem,” presents four empirical findings from persona-topology experiments conducted on Qwen3-4B-Instruct and Mistral-7B-Instruct-v0.2. These findings jointly undermine the assumption that the same direction in representation space picks out the same persona content under prompt-conditioning, gradient-descent fine-tuning, and inference-time steering.

The first empirical wedge is the non-collinearity of prompt-extracted vectors and fine-tune basins, indicating that directions associated with persona content in one regime do not align with those in another. The second is that fictional personas can displace the model along real-anchor directions more strongly than real anchors themselves, suggesting regime-specific sensitivity to persona cues. The third finding is that contradictory-valenced mixtures of personas are biased toward a training-history-determined attractor, revealing regime-dependent stability properties. The fourth is asymmetric compositional algebra under inference-time arithmetic versus fine-tune-time chimera training, highlighting regime-specific compositional behavior.

To address these observations, the authors propose a regime-indexed individuation framework. Under this framework, the identity unit for representational content is a (vehicle, regime) pair, not a vehicle alone. This reframes prior debates—such as those involving Beckmann & Butlin’s three candidate positions, as well as work by Mollo & Millière, Chalmers, and Cerullo—as regime-internal objects rather than competing claims about a single referent.

The paper is structured as a 30-page manuscript with two figures and one table, submitted to the Computation and Language section of arXiv on May 1, 2026. It engages directly with Beckmann & Butlin’s 2026 framework and responds to their arXiv:2604.17031 preprint.