Definition
Plain language
A frozen AI agent that improves on physics puzzles not by retraining, but by keeping a running notebook of principles it distills from its own past attempts, including its failures.
As stated in the literature
Hierarchical Experimentalist Agents — wraps a reactive actor loop in an evolver that contrastively distills reusable skills and anti-patterns (including partial skills mined from failed trajectories) into a capped, ranked bank retrieved into the actor's prompt, all on a frozen model.
Also called: Hierarchical Experimentalist Agents
Why it matters: It matters because it shows an agent can keep getting better by learning from its own attempts without the cost of retraining the model.
For example, after failing a puzzle by launching a ball too hard, the agent writes down 'don't overshoot on steep ramps' and consults that note on later levels.
Heard on the show
“HExA — Hierarchical Experimentalist Agents — takes that inner loop and wraps it in an outer one.”Episode 186 — How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining