Found first: a primary source the press has not covered yet.
A team of 20 researchers has published A-RSI (Adaptive Task Refresh for Zero-Data Multimodal Self-Improvement), a method that detects when a multimodal model has saturated its current training tasks and replaces them automatically. Submitted to arXiv on October 28, 2026, the paper reports an eleven-benchmark multimodal average rising from 57.77 to 61.05 using 800 freshly generated tasks, with no human-written labels at any point in the loop.
What the source says
The method monitors two signals drawn entirely from training logs: an advantage-retention count (the share of prompts still producing nonzero advantages after resampling) and a running accuracy metric. Refresh fires when advantage retention drops below 85% or when accuracy rises by 10 or more percentage points within a window. New tasks are generated 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. The reward structure and refresh logic were formally verified using an SMT solver, though the authors note verification covers only the four properties tested, not full implementation correctness. Authors are affiliated with IIE at the Chinese Academy of Sciences, CUHK, CMU, the University of Würzburg, the Max Planck Institute, and the University of Trento, among other institutions.
Why it matters
Most self-improvement research focuses on which tasks to generate; A-RSI addresses the prior question of when current tasks have stopped being informative. The efficiency gain 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. The benchmarks span math, vision, and chart reasoning, so the average is not inflated by clustering on a narrow task family.