rsi.watch

Yu Liu et al. (20 authors) · 2026-10-28

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

“On the primary eleven-benchmark suite, A-RSI raises the performance average from 57.77 to 61.05 with 800 fresh tasks.”
Who
Yu Liu et al. (20 authors)
Date
2026-10-28
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ imprecise. Every policy is one training run and the paper reports no intervals. Four ladder-control runs with identical first-window data and seed range over 1.25 points at 4096 tokens (sample sd 0.52), and at window 2 and 8192 tokens, two runs whose first two pool files are identical differ by 1.46. The lead of A-RSI over refresh-every-window at window 8 is 0.40. The authors write: "These examples show variability but do not provide replicated uncertainty for each endpoint policy difference." “Each policy arm was trained once.”
  • ↓ risk of bias. The mix of training tasks was set with the evaluation benchmarks in view, and the headline gain is measured on those benchmarks. The paper reports freshness and contamination screens on the 16,000-task reservoir; the initial pool is separate, and the paper says "Two questions in the initial pool were later identified as configuration isomorphs of MathVerse items." The design was chosen over three versions scored on the same eleven benchmarks ("one run per arm"), and the reported method was the "Random control" arm; the paper lists the pre-registered primary endpoint (Frontier - No-Refresh) as not supported: "No-Refresh halted, so there is no reading." “Task-family shares and 19 thematic terms were informed by the evaluation benchmarks; diversity beyond these templates remains unmeasured.”
how we rate evidence
Replication
No rerun found. The paper links no code. We found no rerun by another group. Searched up to 2026-10-30. all claims
Checked
2026-10-30, against the source page
Links
Source Paper

Also on the page

  • Qwen3-VL-8B-Instruct is trained with GRPO on pools of image questions in geometry, charts, tables and counting. Before training, the base model wrote a library of programs that draw each image, write the question and compute its answer; the library then stays fixed. After each window of 100 updates a controller reads the training log. Once a pool has had at least three windows, it replaces the whole pool with 400 unused tasks, drawn at random under fixed family quotas, when the number of prompts with nonzero advantage has fallen by at least 15% or logged accuracy has risen by at least ten points against two windows earlier. Only the model weights are updated.
  • The 57.77 in the quote is the starting checkpoint after 100 GRPO updates of the public Qwen3-VL-8B-Instruct on an initial 500-task pool; the public model scores 56.65. The 61.05 is after eight continuation windows; one more window scores 61.44. On MiMo-VL-7B-SFT the average goes from 59.08 for the public model to 60.71 after eight windows, and is 61.04 at window 6.
  • At window 8 the other refresh policies score 60.64 (refresh every window, 3,200 new tasks), 59.89 (Periodic, 1,200) and 59.59 (Frontier, 800). Keeping the old pool scores 58.35 at window 6 against 60.40, then stops when its health gate fails twice. The authors write that the comparisons with Periodic and Continuous "change both selection and timing, so they do not isolate either mechanism", that with frontier selection held fixed detector-triggered refresh "trails periodic refresh by 0.30 points", and that "the first two realized random-refresh decisions coincide with a three-window schedule".
  • With format-corrected scoring the gain is +3.27, against +3.33 with the original answer matching; all eleven benchmarks rise, and ChartQA goes from 82.84 to 84.96 while truncation moves from 0.60% to 0.56%. On the full suite, Qwen's net contribution from questions finished at both checkpoints is -0.71, against +3.98 from the other strata, and truncation falls from 14.34% to 4.16%. The authors write that fewer truncated answers "do not explain every improvement; it does not determine the overall proportion attributable to reasoning", that "These local observations do not override the full-suite result", and that "the eight-window run was not evaluated at 16k". Against the released MM-Zero v5 weights, A-RSI is ahead by 0.79, which they call "within the estimated training noise".
  • The authors' limits: "Each arm was trained once." "Task-family shares and 19 thematic terms were informed by the evaluation benchmarks; diversity beyond these templates remains unmeasured." "The first refresh point was set using existing logs." Their table of claims lists as not supported "That the detector finds the best refresh time." and "Open-ended recursive self-improvement or unbounded gain (trained to window 8; fixed-pool continuation to window 12)."

In the self-improving AI catalog

  • A-RSI: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that A-RSI cites, oldest first. Under each: where that work sits on the map, then where A-RSI sits (changes in bold), and what A-RSI's own paper says about it. Its whole family on the citation map

Cites STaR 2022, STOP 2023, Self-Rewarding Language Models 2024, Darwin Gödel Machine 2025

← Ledger Report an error