A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement

A-RSI. Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement. By Yu Liu and nineteen others, from the Institute of Information Engineering of the Chinese Academy of Sciences, the Chinese University of Hong Kong, and seven other institutions. Published on the twenty-eighth of October.

Recursive self-improvement asks a model to do two things. It has to acquire useful training experience, and it has to decide when the experience it already has has stopped being useful. Existing methods tend to concentrate on the first question, what experience to generate next, and pay less attention to the second, when fresh experience should come in. There is a further weakness. Many of them build their training targets through self-consistency or majority voting, which means the supervision rests on the model's own uncertain predictions.

A-RSI pulls these two questions apart and answers each separately. Training dynamics decide when to acquire new experience. The answers to new tasks, on the other hand, stay independently checkable, rather than being voted into existence by the model that is learning from them.

For the timing question, A-RSI watches reward variation and accuracy in the training logs, and refreshes its task pool when the learning signal from that pool declines. It does this without ever consulting validation or test scores. For the supervision question, a static library of model-authored programs produces multimodal questions and derives their reference answers. The authors specify the contract that task generation must follow, automatically verify chosen properties of the reward and refresh rules, and separately audit the answers that get generated. Throughout, the training loop uses no external training examples and no human-written answer labels.

The headline result is on the primary suite of eleven benchmarks. There, A-RSI raises the performance average from fifty-seven point seven seven to sixty-one point zero five, using eight hundred fresh tasks.

Two decisions usually tangled together are kept distinct here. When to bring in new tasks is read off the training logs alone, never off held-out scores. What counts as a correct answer comes from programs whose outputs can be checked, not from the model agreeing with itself. That separation is the whole design, and it lets a self-improving loop run with zero external data while keeping both its schedule and its targets accountable.