LLM individuation: Cheng's regime-indexed individuation case

Shuaizhi Cheng submitted a 30-page paper to arXiv on 1 May 2026 arguing that persona representations in large language models do not persist unchanged across prompt-conditioning, fine-tuning, and inference-time steering. The paper presents experiments on Qwen3-4B-Instruct and Mistral-7B-Instruct-v0.2 and proposes "regime-indexed individuation," where the identity unit for representational content is a (vehicle, regime) pair rather than the vehicle alone.

What does Cheng claim about persona vectors and individuation?

Cheng argues that a core assumption in persona-vectors literature, the cross-regime co-reference claim that the same direction picks out the same content across regimes, is unargued and empirically undermined. The paper offers four empirical wedges from persona-topology experiments: non-collinearity between prompt-extracted vectors and fine-tune basins; fictional personas displacing models more along real-anchor directions than those anchors themselves; mixtures with contradictory valence biased toward a training-history-determined attractor; and asymmetric compositional algebra when comparing inference-time arithmetic to fine-tune-time chimera training.

How did the experiments demonstrate regime dependence?

Cheng ran persona-topology experiments specifically on Qwen3-4B-Instruct and Mistral-7B-Instruct-v0.2 to probe whether the same geometric directions correspond to the same representational content across regimes. The paper reports empirical phenomena including non-collinearity of prompt-extracted vectors and fine-tune basins, displacement effects where fictional personas move the model along real-anchor directions more strongly than real anchors, biased mixtures that converge toward an attractor set by training history, and asymmetric algebraic composition between inference-time arithmetic and fine-tune-time chimera training. The manuscript is 30 pages long and includes 2 figures and 1 table, and it is framed as a reply to Beckmann & Butlin (arXiv:2604.17031).

How does Cheng formalize an alternative?

Cheng proposes regime-indexed individuation: the identity of representational content is not a vehicle alone but a (vehicle, regime) pair. Under this framework, Beckmann & Butlin's three candidate positions do not compete for the same referent; instead, each describes a regime-internal object. Cheng extends the same diagnosis to analyses from Mollo & Millière, Chalmers, and Cerullo, arguing they face the same cross-regime co-reference vulnerability.

Why it matters

If persona directions are regime-dependent, then methods that equate prompt-induced directions, fine-tuning changes, and inference-time steering risk conflating different internal objects. That matters for theoretical claims about model semantics and for practical work that transfers interventions or interpretations from one regime to another. Cheng's regime-indexed move forces researchers to track how an intervention's referent may shift when the vehicle remains nominally the same but the regime changes.

What to watch

Look for community responses to Cheng's reply to Beckmann & Butlin and for empirical replication on other models and scales. Confirming regime-indexed differences beyond Qwen3-4B-Instruct and Mistral-7B-Instruct-v0.2 would shift interpretation practices for persona vectors and for cross-regime intervention methods.

References and concrete facts in this brief are drawn from Shuaizhi Cheng's arXiv submission "Persona Without Substrate: Regime-Dependence and the LLM Individuation Problem," submitted 1 May 2026 (30 pages, 2 figures, 1 table), which presents experiments on Qwen3-4B-Instruct and Mistral-7B-Instruct-v0.2 and is framed as a reply to Beckmann & Butlin (arXiv:2604.17031).