PHYSICS OF AIThe physical principles of intelligence.

Measured IdeaSearchFitter three-dimensional potential landscape

Detailed balance in large language model-driven agents

We ask whether state-to-state transitions in LLM-driven agents admit an effective potential. Coarse-graining many token trajectories into semantic states lets us test measured forward–reverse probability ratios against a potential landscape across tasks and model snapshots.

Linear-response curves for information susceptibility in LLM agents

Information susceptibility in LLM agents

We treat a fixed LLM as a linear-response channel: world knowledge gives a head start, while hallucination limits the marginal return as the budget grows.

Schematic phase diagram: a trained inference flow settles in the central working phase, while untrained flows drift toward noise and collapse

A phase diagram for training and inference in language models

Two collective coordinates place every layer of a language model on one phase diagram. Trained models settle in a central working phase, while most untrained networks drift toward collapse or noise; training pulls the inference flow into the working phase, which we interpret as a phase transition.