A-RSI lifts multimodal average from 57.77 to 61.05 unaided

A-RSI, an adaptive task-refresh method for multimodal self-improvement, raised a model's average score on an eleven-benchmark suite from 57.77 to 61.05 using 800 freshly generated tasks and no human-labeled data, according to a paper posted to arXiv.

The approach pulls apart two problems most self-training systems entangle. Rather than scoring on validation sets to decide when to swap in new material, A-RSI watches its own training logs: when reward variation and accuracy flatten, the task pool is refreshed. "A-RSI monitors reward variation and accuracy in training logs to refresh a task pool when its learning signal declines, without consulting validation or test scores," the authors write.

Answers come from a separate verification path. A "static library of model-authored programs produces multimodal questions and derives their reference answers," the abstract says, which the authors contrast with the majority-voting and self-consistency schemes they argue leave "supervision depend on the model's own uncertain predictions." The training loop "uses no external training examples or human-written answer labels."

The abstract publishes only the eleven-benchmark average, not per-benchmark numbers or the identity of the base model being improved.