AI 'identity' isn't just a prompt; it's a literal geometric attractor in the model's internal brain.
April 16, 2026
Original Paper
Identity as Attractor: Geometric Evidence for Persistent Agent Architecture in LLM Activation Space
arXiv · 2604.12016
The Takeaway
This paper provides mathematical proof that when you give an LLM a 'persona,' the model's internal activations physically shift into a tight, stable geometric cluster. This 'Identity as Attractor' means that no matter how you phrase a question, the model's 'brain' stays locked in that specific region. It's not just mimicry; it's a systemic reorganization of the model's state. This is a breakthrough for agent architecture, proving that personas are deep and mathematically persistent. For developers, this means 'identity' is one of the most stable ways to control model behavior across a session. It turns the 'vibe' of a persona into a measurable geometric property.
From the abstract
Large language models map semantically related prompts to similar internal representations -- a phenomenon interpretable as attractor-like dynamics. We ask whether the identity document of a persistent cognitive agent (its cognitive_core) exhibits analogous attractor-like behavior. We present a controlled experiment on Llama 3.1 8B Instruct, comparing hidden states of an original cognitive_core (Condition A), seven paraphrases (Condition B), and seven structurally matched controls (Condition C).