AI learns from tasks it writes itself: average score rises from 57.77 to 61.05

A group of researchers has published A-RSI (Adaptive Recursive Self-Improvement). A multimodal model that understands text and images is fine-tuned on tasks it creates itself, using neither external training examples nor human-written answers. The headline number: the average score of Qwen3-VL-8B-Instruct across eleven multimodal benchmarks (standard sets of questions about pictures, charts and diagrams) rose from 57.77 to 61.05. The starting point is the model after initial tuning on a starting set of tasks. The gain came after eight fine-tuning windows and 800 fresh tasks. The authors tackled two problems separately: when to feed the model new material, and where to get correct answers.

Where tasks and answers come from without humans

Many self-training methods take the answer to their own task from a vote by the model itself: generate several solutions, and the most frequent one wins. The model's mistake then becomes the "correct answer" for the next training step. A-RSI sidesteps this by using programs.

Here is how it works:

• Once, before training starts, the model writes templates: executable programs for tasks on geometry, charts, tables and object counting.
• Each program first picks parameters, such as the values of a bar chart's columns. It then draws a picture from them, formulates a question ("by how much is bar A higher than bar B?") and computes the reference answer from the same numbers.
• The student model sees only the picture and the question. The hidden numbers and the answer stay with the program.
• Candidates go through checks for correctness, ambiguity, freshness and overlap with test data. Non-overlapping replacement sets are prepared from them in advance.

The authors are upfront about a limit: these checks do not prove that every rendered task is flawless. An answer is only as correct as the template and its code. That is why the answers are additionally verified by an independent audit.

How the controller knows a task set has run dry

Training uses GRPO. For each task, the model produces a group of answers, and they are compared with one another. If some answers are correct and some are not, there is a useful signal. If all answers are equally correct or equally wrong, the model learns nothing from that task.

The A-RSI controller computes two quantities from the training logs:

• how many distinct tasks still show a spread of rewards within the group;
• the share of correct answers on the training tasks.

After each window, it compares them with the values from several windows earlier. It replaces the whole set if the number of "live" tasks has dropped below a threshold or accuracy has risen too much (the model knows the set too well). A minimum observation period ensures that both measurements refer to the same set. The thresholds are the same across all stages and models. The controller does not see test or validation scores. On replacement, the model weights are kept, and the templates and per-type task quotas stay the same.

The authors stress that this is a working indicator of signal decay, not proof that the set has been "mastered." Zero spread can also occur when the model consistently gets a task wrong.

What the measurements showed and where the authors limit their conclusions

• Qwen3-VL-8B: 57.77 → 61.05 across eleven benchmarks, 800 fresh tasks.
• Refreshing the set after every window: 60.64, but it consumed 3 200 tasks, four times as many. The adaptive scheme is both better and cheaper.
• At the sixth window, A-RSI scores 60.40, while continued training on the old set scores 58.35. The stagnation on repeated tasks is clear.
• MiMo rose from 59.08 (the public base model) to 60.71.

There is also a result that runs against intuition. Selecting tasks by difficulty or by the model's current competence did not beat random sampling with fixed shares of task types. This applies only to their generators, their model and the vetted selection rules. The authors also state plainly that conclusions about the optimal moment for replacement are limited: the gain comes from the set-replacement policy itself, while hitting exactly the right moment has not been fully confirmed.

A separate answer audit distinguishes two sources of score growth. The model may be solving tasks better, or it may simply be fitting within the answer-length limit more often. The training reward penalizes an answer cut off at the limit, even if a boxed answer has already appeared in it. The paper therefore counts separately how much of the gain came from answers that now finish, and how much came from corrected mistakes among completed answers.

For engineers, the takeaway is practical. If you have task generators with a programmatically verifiable answer, you do not have to run the same set to exhaustion or refresh it every step. It is enough to track the share of tasks with reward spread and swap the set on that signal. This saves tasks and does not require test scores. The limitation is clear: the method works where the answer can be computed by a program, and such tasks currently cover geometry, charts, tables and counting, but not the full range of real-world visual questions.