Got the abstract; now pulling the full text (via arXiv's HTML version) so the review covers methods and results, not just the summary.

Review (ELI12): "Recursive Self-Improvement through Multi-Agent Self-Supervision"

Lee, Xu, Seely, Lee, Sojoudi, Zaharia, Tang. arXiv, 8 Oct 2026.

What they're trying to do

Picture a student who wants to get better at writing science reports, but no teacher is available to grade them. The student has to grade their own work. That usually goes badly, because you tend to miss your own mistakes.

The paper's idea is to have the student split into a team. One copy acts as the manager. Others act as helpers, like "the data person" or "the person who checks the code." Every team member is the same AI, Qwen3.6-27B, playing different roles.

Here's how the process works:

1. Try out team setups. The AI makes up a way to organize the team: who does what, and what each helper reports back. It runs the team and compares the result to the best setup so far. If the new one wins, it stays. This repeats many times, like breeding better and better recipes.
2. Study your best work. The AI runs the winning setup 18 times, ranks its own attempts, and trains on the top 15.
3. Repeat. The trained AI is now a better worker and a better judge, so the next round should go better.

They call this MASS and ran it for two rounds.

What they found

• On their own made-up research tasks, it clearly improved. Two outside judges (GPT-5.5 and Claude Opus 4.8) preferred the round-2 AI over the original 82% of the time on the training tasks. On the 3 held-out test tasks, they preferred it 70% of the time.
• The AI learned to delegate without being told to. With a plain prompt, the original AI handed work to helpers in 1 out of 110 tries. After round 2 it did so in 87 out of 110. It actually picked up the habit.
• It got better at judging. Its own verdicts matched the outside judges 73% of the time at the start and 93% after round 2.
• Team transcripts make better study material. A student AI trained on team transcripts beat one trained on single-AI transcripts (68% vs 64%), using about 31% less training text.

Where I'm skeptical

1. The public test results are weak. The headline "1.2 to 1.6 times better" is measured per output token: score divided by how much the AI wrote. A model can look better on that measure partly by writing less. The raw scores:

   Benchmark	Before	After 2 rounds	My read
   MLR-Bench	1.58	2.40	Real gain
   ScienceAgentBench	28.8	31.7	Small, error bars overlap
   AstaBench	0.053	0.067 (±0.017)	Mostly noise
   DSBench	0.473	0.482	Flat
   Terminal-Bench 2.0	37.5	35.2	Worse
   SWE-bench Verified	62.6	61.1	Worse

   The coding scores went down. The abstract only mentions the four research benchmarks, and I didn't see the coding drops discussed in the main text.

2. The test is very small. The strongest results come from 3 held-out tasks, and all the training and test tasks are private, so nobody else can rerun them.

3. One judge has a conflict of interest. GPT-5.5 wrote the made-up tasks and is also one of the two judges.

4. "Recursive" rests on two rounds. Round 2 did gain more than round 1, which is encouraging. But two rounds can't show whether this keeps working or levels off.

5. The comparison isn't quite even. For the single-AI students it was compared against, they picked training snapshots using the outside judges, which the main method doesn't get to do. They do admit this openly.

6. We don't know the full cost. They report training time (about 19 hours on 6 H100s per round), but not the cost of all the team runs and judging, which is probably the bigger bill.

Bottom line

It's a smart idea with one convincing result: the AI really learns to organize a team of itself, and team transcripts are better study material than solo ones. But the title promises "recursive self-improvement," and the evidence is two rounds, 3 private test tasks, and public results that are mixed and slightly worse on coding.

Verdict: interesting proof of concept, oversold in the framing.

Caveat: I couldn't render the PDF here, so I read the arXiv HTML version through a summarizing tool. That covered the main paper but not the back half of the appendices, so a few numbers could be slightly off.