Research note
Research note: Sakana AI and UC Berkeley researchers publish MASS, a method for a model to train itself using its own multi-agent workflows
How we checked what the MASS paper claims and where its results fall short.
What we set out to check
What MASS is, which model it was tested on, what results the researchers report, and whether any benchmarks got worse.
What we found
- The project page and arXiv listing agree on the method, the model (Qwen3.6-27B) and the 1.2–1.6x range on four research benchmarks.
- The project page also shows two coding benchmarks where the trained model scored slightly below the base model.
- The page quotes several limitations, including noisy workflow search and uneven transfer.
Numbers that differ
AI Weekly lists fewer authors than the project page and arXiv, which name seven. We use the full list from the paper. Benchmark figures matched across sources.
Labels we chose and why
- Confirmed for the publication itself, read on Sakana's page and arXiv.
- Claim for every result, since they are the researchers' own measurements and have not been independently reproduced.
What we could not verify
- Independent replication of the results.
- Whether the code at the linked repository reproduces the paper's numbers.
What we searched
- Sakana AI MASS multi-agent self-supervision
- arXiv 2610.12176
- Sakana MASS Qwen3.6-27B benchmark
Every source we read
| Source | Used? | Why |
|---|---|---|
| Sakana AI project page | USED | Method, authors, full benchmark table, quoted limitations and code link. |
| arXiv 2610.12176 | USED | Title, authors, submission date and abstract summary of results. |
| AI Weekly | USED | Secondary summary; used only for the authors' blind-spot warning and to cross-check numbers. |