The continual improvement of large language models increasingly relies on synthetic experience, yet for interactive agents the data must be executable, verifiable, and provide learning value for the current task. Existing approaches incorporate environment feedback, data validation, or learner failures, but they still suffer from a feedback‑to‑decision gap: observed weaknesses do not directly specify how future synthesis should change, nor whether they can be represented within the current synthesis space.
To bridge this gap we introduce EvoSyn, a learning‑feedback‑driven multi‑agent framework. The framework organizes synthesis around a revisable state that defines which interaction factors can be deliberately generated and where synthesis effort should be allocated. Role‑specialized agents translate learner feedback into synthesis revisions; a Designer allocates synthesis directions and a Planner grounds them into executable experience. This enables within‑space adaptation toward weak or under‑covered regions and expands the synthesis space by promoting newly discovered factors into explicit generation controls.
Experiments on $$\tau^2\text{-Bench}$$ and CAR‑Bench show that EvoSyn improves average task success by 13.3 points over the strongest baseline, while producing higher‑quality, structurally broader synthetic experience and uncovering controllable factors missed without learner feedback.
Review