Concept study / Information & adaptationConceptual illustration. Shapes, colors, and motion are not measured data.
An adaptive material is more than an object that moves when pushed. In a learning architecture, its adjustable properties change through training so that future inputs produce improved outputs for a defined task. Stern and collaborators developed a framework for local learning rules in physical networks. Dillavou and collaborators demonstrated decentralized learning in an electrical network of adjustable resistors. These are strong examples for Hill’s interest in geometry and retained response. Their scope is an engineered system with a specified training procedure. They are not direct evidence that all matter learns or that the universe shares the same update rule.
Where geometry enters
A network’s topology and adjustable elements constrain the responses it can produce. In a mechanical network, the connections and stiffnesses affect force transmission; in an electrical network, conductances and connectivity affect voltage responses. We propose treating geometry and material parameters as separate experimental variables. Otherwise an improvement attributed to shape may actually come from changing a component’s properties or the amount of supplied power. A useful model records what was adjusted during training and what remained fixed throughout the comparison.
A benchmark researchers can reproduce
The proposed first benchmark is a small network learning a simple input–output mapping. Publish the initial network, update rule, target outputs, random seed, and loss curve. Compare with a frozen network and a random-update control. Report energy use and convergence time separately from prediction error. Readers should be able to inspect cases where training fails or where the architecture cannot represent the target. This makes the tool useful as a research exercise before any claim is made about new materials or universal principles.
What a positive result would mean
If geometry changes produce a repeatable advantage under matched conditions, the result supports an architectural claim for that model and task. Generality would require testing different tasks, system sizes, and physical realizations. We propose asking whether a dimensionless relationship predicts those changes without being refitted each time. That is a plausible bridge toward Hill’s cross-scale question. It remains a large step from such a bridge to a cosmological theory, and the evidence map should show the intermediate tests still missing.
The mathematical lens
Illustrative learning update; a physical implementation must explain how local processes estimate the change.
Turn it into a question.
- Question
- Which structural changes cause better held-out performance?
- Method
- Vary topology and material parameters independently with fixed-resource controls.
- What would count against it?
- No repeatable advantage after accounting for parameters and resources weakens the geometry-specific claim.
Engineered physical learning is real; universality is a further hypothesis.
Follow the original work.
These sources support the mechanisms and limits discussed here. They do not endorse Hill’s framework. Notes distinguish research proposals, reviews, and observations.
Continue the investigation
Edition 02 · 18 October 2026 · AI-assisted synthesis. See editorial method.