A-RSI trains vision models on self-made tasks, gaining 3.3 points

A-RSI, a method for multimodal self-improvement, trains vision-language models. It decides when to swap in fresh practice problems by watching its own training logs, not by checking test scores. In the paper, posted on October 28, 2026, the authors report that it lifted Qwen3-VL-8B's average score across eleven benchmarks from 57.77 to 61.05 with no outside training data. It was the second most upvoted paper on Hugging Face daily papers at the time of writing, with 148 upvotes.

The problem in multimodal self-improvement: models run out of things to learn

Zero-data self-improvement means a model trains on problems it makes itself, with no human-written examples or labels. The trouble is that a fixed set of practice problems goes stale. After a while the model gets most of them right, or all of the attempts at a problem look alike, and the training signal fades.

The authors split the fix into two questions. When should the model get new problems? And where do the new problems come from, with answers that don't depend on the model's own guesses? Many earlier methods label problems by majority vote among the model's own answers. That ties the supervision to the model's uncertainty.

What the team did

The base model first writes small programs, each a template for one kind of task: geometry, charts, tables and counting. A program picks random parameters, draws an image, writes a question and computes the correct answer, all from the same settings. So the answer is exact, not voted on. The library of programs is fixed before training starts.

Training uses GRPO, a reinforcement learning method that samples several answers per problem and rewards the better ones. It only teaches the model that solves the problems. A problem teaches something only when its sampled answers differ in reward, so the method counts how many distinct problems still show that variation in each window of 100 updates.

When that count drops, or average accuracy climbs past a threshold, the whole pool is replaced. The paper cites an 85% retention threshold for the count and a ten-point accuracy threshold. The new pool is 400 unused tasks, picked at random within fixed quotas for each task family. Nothing from the validation or test sets is used to decide.

The results

The main model is Qwen3-VL-8B-Instruct. Scores are the equal-weight average over eleven test sets, including MathVerse, MathVision, MMMU, MMMU-Pro, ChartQA and PhyX. The authors report:

• The starting model scores 56.65. After eight windows, A-RSI reaches 61.05 and a ninth window gives 61.44.
• The 57.77 to 61.05 figure in the abstract starts from a checkpoint trained for 100 updates on the initial pool.
• Training on the old pool without refreshing reaches 58.35 at the sixth window. A-RSI scores 60.40 at that point, a gap of 2.05 points.
• Replacing the pool every window used 3,200 new tasks and scored 60.64. A-RSI used 800 and scored 61.05.
• On a second model, MiMo-VL-7B-SFT, the average rises from 59.08 to 60.71. That is a smaller gain, and the last checkpoint drops back to 60.35.

The paper also compares against a released checkpoint from MM-Zero, an earlier zero-data method. With an 8,192-token answer limit, A-RSI is 3.25 points above its base model, against 2.45 for MM-Zero.

One finding runs against intuition. Choosing replacement tasks that look most useful did worse than choosing them at random. Picking tasks the model gets wrong scored 2.63 points below the control, and a learnable-difficulty band scored 1.90 below. The authors say this holds for their task generators, model and rules, and they don't claim it is general.

How much to trust it

The authors list their own limits, and several matter.

• Each refresh policy is a single training run. There are no repeated runs to show how much results vary by chance.
• The paper does not show that the log-based trigger is what helps. A variant that used the trigger with a different selection rule scored 59.59, below a fixed-schedule variant at 59.89. The authors say the timing advantage is not established.
• Answers rely on the programs being correct. Independent solvers agreed with all 1,000 audited answers, but the answer checker failed 220 of 16,000 adversarial test cases, mostly on multiple-choice questions where an ambiguous answer was accepted. The fix was not used in the reported runs.
• The tasks cover four families. How far this carries to other kinds of reasoning is untested here.
• I read the arXiv text but not the appendices, and I did not find a link to code, weights or data in the pages I read.

Why it matters

Gains of three to four points on an average are modest, and the models are small. The useful idea is narrower: you can generate training problems whose answers are computed, not guessed, and you can tell from training logs when a pool has stopped paying off. If that holds up in repeated runs, it makes self-training cheaper to run without a labeled validation set.

For now it is one paper with single runs. Watch for a code release and for replications on other model families. Related work on our site covers self-improvement in a different setting.