A-RSI lifts 11-benchmark average 57.77 to 61.05 with no labels

A-RSI comes from a team of 20 researchers affiliated with institutions including the Chinese Academy of Sciences' IIE, CUHK, CMU, the University of Würzburg, the Max Planck Institute, and the University of Trento. The work was submitted to arXiv on October 28, 2026.

Most self-improvement research focuses on which tasks to generate, but A-RSI addresses the prior question of when current tasks have stopped being informative. The method generates new tasks from a static library of model-written programs that produce images and independently verifiable reference answers, so correctness checking does not depend on model self-consistency. This design choice may be what allows the loop to operate without human-written labels at any point.

The efficiency comparison in the paper is concrete: a continuous every-window refresh baseline reaches 60.64 consuming 3,200 tasks, while A-RSI reaches 61.05 using 800, roughly one-quarter the task budget for a higher final score. The zero-label constraint matters because it removes the bottleneck that makes most recursive self-improvement pipelines impractical at scale, according to the authors. The benchmarks span math, vision, and chart reasoning, so the average is not inflated by clustering on a narrow task family. The authors note that their SMT solver verification covers only the four properties tested, not full implementation correctness, which may be a limitation for teams considering deployment.