This framework explores concept formation in large language models, suggesting new dynamics and implications.
⚠ This is an early draft (v0.1) and is actively under development via multi-model collaborative refinement (3-body: GPT / Gemini / Claude). The document is uploaded for timestamp preservation and priority establishment only. Formal submission to a peer-reviewed venue will follow after structural consolidation. We propose a Phase Potential framework for understanding why large language models (LLMs) encode specific abstract concepts — biases, moral binaries, personality traits, value systems — as identifiable phase structures in their latent representation spaces. Building on the logarithmic cost scaling (W ~ ln n) established in our prior observation-theoretic work (see Related Records), we model concept formation as cost-driven attractor dynamics shaped by unresolved representational tensions in human-generated training data. This work is motivated by and complementary to Radhakrishnan et al. (Science, 2026), "Toward universal steering and monitoring of AI models," which demonstrated extraction and steering of 500+ concepts via Recursive Feature Machines (RFM). While their work addresses what concepts exist and how to steer them, this framework addresses why those particular phase structures form and predicts cross-model universality of attractor geometry. Key contributions (draft stage): Phase potential formulation linking training data tensions to latent space attractor basins Logarithmic depth scaling of attractors from observation-theoretic cost structure Cross-model universality hypothesis for architecturally distinct LLMs Reframing of steering vectors as external field perturbations on phase potential Five experimentally testable predictions Related records: CH Paper: "Continuum–Discrete Paradoxes, Base–Topological Waveframes..." (APS submitted but reject by no category) Involutive Boundary & Resonance Notes LINKS THAT ARE RECEIVABLE OF THE RESULTS OF THE PROCESS Supporting materials on Hugging Face repository Status: Draft v0.1. Multi-model refinement in progress. Updates will be versioned on this record.
No takes yet. Share an insight, caveat, or question.
Kimminsu No-Pattern Engine (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: