When agents continuously generate and revise skills, the mechanism that discovers and refines those skills becomes a learnable object itself. Task‑skills act directly on execution, while meta‑skills govern how agents discover and improve future skills, and their value emerges through the subsequent search processes they induce. Existing methods improve meta‑skills by observing raw skill‑search trajectories and branch outcomes, but branch performance entangles the initial discovery state with the meta‑skill revision that generated the search, making it hard to pinpoint what a particular revision actually changed. Consequently, updates tend to exploit favorable states rather than truly improve the process. We introduce HMED (Hindsight Meta‑Experience Distillation), which revisits the completed event that gave rise to a revision and re‑executes both the incumbent and revised meta‑skills from the same restored discovery state, allowing the effects of the revision to be observed under shared conditions. Each comparison is distilled into a meta‑experience record that can be reused by future updates, so even revisions that are not ultimately kept still provide a learning signal. Across three interactive agent benchmarks and both open‑source and closed‑source models, HMED consistently outperforms strong baselines, boosting skill discovery performance and shifting meta‑skill learning from branch outcomes toward the consequences of changing the improvement process.
Review