Recursive self‑improvement demands that models not only acquire useful training experience but also decide when that experience ceases to be beneficial. Prior work often concentrates on what experience to generate next while paying little attention to when fresh experience should be introduced; many also build training targets via self‑consistency or majority voting, making supervision depend on the model's own uncertain predictions.
A‑RSI separates these concerns: training dynamics determine when to acquire new experience, and task answers remain independently verifiable. The system monitors reward variation and accuracy in training logs, refreshing the task pool whenever the learning signal declines, without consulting validation or test scores.
A static library of model‑authored programs produces multimodal questions together with reference answers. We specify a task‑generation contract, automatically verify the chosen properties of reward and refresh rules, and audit the generated answers separately. The training loop uses no external examples or human‑written answer labels.
On the primary eleven‑benchmark suite, A‑RSI raises the average performance from 57.77 to 61.05 after adding 800 fresh tasks.
Review