On May 1, 2026, Shuaizhi Cheng presented a groundbreaking paper titled Persona Without Substrate: Regime-Dependence and the LLM Individuation Problem. This research examines the LLM individuation problem within the context of artificial intelligence, challenging existing assumptions about how language models identify and respond to prompts.
Challenging Existing Frameworks in LLMs
Cheng critiques the ontological framework proposed by Beckmann & Butlin, which suggests that language models maintain consistent co-reference across different regimes. Cheng argues that this framework inherits an untested assumption from persona-vectors literature, which may not hold true in practice.
The paper presents four empirical findings from experiments conducted on models such as Qwen3-4B-Instruct and Mistral-7B-Instruct-v0.2. These findings reveal significant discrepancies in how prompt-extracted vectors relate to fine-tuning outcomes. For instance, Cheng notes that fictional personas can influence the model's responses more dramatically than real anchors.
Empirical Findings on LLM Behavior
Cheng's experiments reveal several key insights:
- Non-collinearity of prompt-extracted vectors and fine-tune basins.
- Fictional personas can displace the model along real-anchor directions.
- Contradictory-valenced mixtures are biased toward a training-history-determined attractor.
- Asymmetric compositional algebra under inference-time arithmetic differs from fine-tune-time chimera training.
These findings collectively suggest that the previous assumptions about LLM behavior need reevaluation.
Proposing a New Framework for LLM Individuation
Cheng introduces the concept of regime-indexed individuation, positing that the identity unit for representational content should be viewed as a (vehicle, regime) pair rather than as a vehicle alone. This approach allows for a more nuanced understanding of how language models operate within different contexts.
According to Cheng, Beckmann & Butlin's positions represent different regime-internal objects rather than competing for the same referent. This perspective can also be applied to the works of Mollo & Millière, Chalmers, and Cerullo, suggesting a need for a broader view of LLM individuation.
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by arXiv NLP. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
