Theory-mediated learning in adaptive-system experiments
Status: Exploratory Commonplace workshop. The theory-mediated additions below are our proposals, not claims made or tested by the cited systems.
Three current systems expose different parts of a learning or improvement loop. HCL proposes, evaluates, and retains harness changes. SPADE generates and filters executable learning environments. Exo provides an inspectable, mutable substrate in which accepted self-changes can remain operative. Each system provides a setting in which an experiment could test the same missing treatment: place a scoped, criticizable theory of system behavior between evidence and the decision that uses it.
Commonplace describes this mechanism as theory-mediated learning. An LLM constructs or retrieves a theory, then uses the theory's consequences to guide diagnosis, candidate search, candidate selection, evidence acquisition, or outcome interpretation. Initial experiments can construct a working theory within each episode. Later experiments can ask whether retaining and revising theories improves subsequent episodes. This workshop turns the proposal into comparisons that can fail. It does not assume that an explicit theory will be correct, useful, or cheaper than direct search and full evaluation.
Reading map
Core proposal
- Theory-mediated improvement loop — where theory enters search, selection, evaluation, and retention.
- Selective-evaluation model — the costly-evidence application.
- Experiment design — comparisons of direct, on-the-spot-theory, and retained-theory treatments.
Experimental connections
- HCL: reading and author invitation.
- SPADE: source analysis and author invitation.
- Exo: case, evidence ledger, and author invitation.
Closing the workshop
Close when the proposed comparisons are precise enough to hand off or reject, and any durable conclusions have been promoted to the library or explicitly declined.
Complete file listing (generated at build time)