Recursive self-improvement loops usually stall on a question nobody optimizes for: when has the current experience run dry? A-RSI, posted to arXiv on October 28 by Xiaomi Research with the University of Trento and others, offers an engineering-grade answer — watch the model's own training log, and swap the whole task pool when the learning signal fades. No external labels anywhere in the loop. The paper reached #2 on Hugging Face Daily Papers, with 155 upvotes at writing time.

Two old problems

Most self-improvement work focuses on what experience to generate next. A-RSI highlights two gaps: supervision is often built from self-consistency or majority voting — the model grading its own uncertain guesses — and almost nothing decides when fresh experience should be introduced, so training spins on a spent pool.

Training log as referee, programs as question writers

A-RSI separates the two. Refresh timing: it monitors reward variation and accuracy in training logs only; when signal retention drops below an 85% threshold, the entire 400-task pool is replaced, drawn under fixed family quotas (194 geometry, 131 chart, 54 table, 21 counting). No validation or test scores are consulted. Supervision: a static library of model-authored programs generates multimodal questions and computes reference answers that are independently checkable — no voting. The refresh rule itself is formally verified across 321,605 threshold and initialization cases.

800 fresh tasks, +3.27 points

On Qwen3-VL-8B-Instruct, eight continuation windows and 800 fresh tasks lift the Macro11 average from 57.77 to 61.05, with all 22,042 test items improving (+1.19 to +8.33 per benchmark). The controls tell the story: continuing on the old pool plateaus at 58.35 by window 6; refreshing every window burns 3,200 tasks to reach 60.64. A-RSI beats that by 0.40 with a quarter of the tasks — 4.09 points gained per thousand tasks versus 0.90. On Xiaomi's MiMo-VL-7B-SFT, scores rise from 59.08 to 60.71, and the detector fires at window 3, as on Qwen.

The honest audit is the best part

The authors puncture part of their own gain: truncation falls from 14.34% to 4.16%, so a chunk of the improvement is the model learning to finish within budget rather than genuinely stronger visual reasoning — format-corrected rescoring still preserves +3.27. An independent Qwen2.5-14B-Instruct judge puts accuracy at 59.54 versus 54.83 at the start, and doubling the output budget from 4096 to 8192 tokens leaves a 3.25-point gain over base, ahead of the prior zero-data method MM-Zero v5 at 2.45. Each continuation window runs 100 GRPO updates. They also report that difficulty, controller, and rollout-based selectors fail to beat random draws, and that single-run margins do not establish a timing advantage. The project page lists code as coming soon.

The bottleneck for zero-data self-improvement is moving from generating tasks to knowing when to refresh and keeping supervision independent. For teams running self-improvement loops on their own stacks, the reusable asset is the sequence — formal verification, independent audits, stated limits — not the 61.05 number.

Paper: https://arxiv.org/abs/2610.38568 ; Hugging Face: https://huggingface.co/papers/2610.38568