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

The Persona Trap: Why LLM Individuation Remains an Unsolved Problem

A new preprint from Beckmann & Butlin (2026) exposes a critical weakness in how the AI community conceptualizes and identifies distinct personas within large language models. Their work, posted on arXiv, argues that the prevailing "persona-vector" approach—which treats specific activation directions as stable markers of identity—rests on an unjustified assumption: that the same vector direction picks out the same semantic content across different model regimes (e.g., different training runs, checkpoints, or fine-tuning states).

This matters because the LLM individuation problem—how to determine whether two outputs come from the same "person" or distinct agents—has become increasingly urgent. As models are deployed with persistent memory, role-playing capabilities, and multi-agent architectures, practitioners need reliable methods to track identity. The persona-vector literature promised a simple solution: find a direction in activation space that corresponds to a consistent persona, then use that direction as a fingerprint. Beckmann & Butlin show this promise is premature.

The core issue is what they call "cross-regime co-reference failure." A vector that encodes "helpful assistant" in one model checkpoint may encode something entirely different after fine-tuning, or even under different prompting conditions. The substrate—the model weights and architecture—shifts, and the persona supposedly anchored to that substrate shifts with it. Without a substrate-independent individuation criterion, we cannot guarantee that "Claude-v1 helpful" and "Claude-v2 helpful" refer to the same agent, even if the vector directions appear similar.

Why This Matters for AI Practitioners

For developers building on LLMs, this has immediate practical implications:

• Persistent identity is an illusion. If you rely on persona vectors to maintain consistent character across sessions or model updates, you may be creating the appearance of continuity without the substance. Users interacting with what they believe is a single agent may actually be engaging with a succession of different entities.

• Safety alignment may be brittle. If alignment vectors (e.g., "harmlessness") suffer from the same cross-regime co-reference problem, then safety guarantees that hold in one model version may not transfer to another. This is particularly concerning for models that undergo continuous fine-tuning or are deployed in federated settings.

• Evaluation metrics need rethinking. Benchmarks that measure persona consistency across contexts may be measuring superficial vector similarity rather than genuine identity persistence. The field needs new metrics that account for regime dependence.

Key Takeaways

• Beckmann & Butlin demonstrate that persona-vector approaches to LLM individuation rely on an unproven assumption that vector directions maintain stable semantic reference across different model regimes.
• This undermines claims about persistent agent identity in LLMs, with direct implications for deployment scenarios requiring long-term user relationships or multi-agent coordination.
• Safety and alignment researchers should treat vector-based identity markers as provisional, not foundational, and develop regime-robust individuation criteria.
• Practitioners should avoid anthropomorphizing persona consistency and instead build systems that explicitly handle identity transitions, rather than assuming them away.