Sakana AI reports 1.2–1.6x efficiency gains after two self-improvement cycles with MASS
After two recursive self-improvement cycles, the researchers behind MASS claim a 1.2–1.6x efficiency gain on four research benchmarks, a regression on coding tasks — and themselves point to the risk that the model's blind spots get reinforced.
Sakana AI published the paper "Recursive Self-Improvement through Multi-Agent Self-Supervision" on October 12, 2026, describing the method MASS (Multi-Agent Self-Supervision). The idea is seemingly simple: one and the same open-weights model — Qwen3.6-27B — solves tasks through virtual subagent layers, searches for better workflows, and then uses its own judgment to pick out the best runs, which the model is then fine-tuned on.
"One shared model solves tasks through a layer of virtual subagents, searches for better multi-agent workflows, and then uses its own evaluations to select the best," Sakana AI writes on its blog. "Training on the best layer runs distills the layer's collective experience back into the shared model."
The method is called recursive self-improvement (RSI) because the process is repeated: the model improves, and the improved model is used to improve itself again. After two such cycles, Sakana reports that score per output token reached 1.2 to 1.6 times the base model's level on four research benchmarks. But the same publication warns that when the model is its own evaluator, shared blind spots can allow flawed results to be accepted as progress and reinforced through training.
How MASS works mechanically
MASS consists of an inner and an outer loop. In the inner one, the model solves tasks through a layer of virtual subagents. According to The Decoder, the workflows are in practice prompt text that defines the subagent roles through so-called "contracts" and "hops" — that is, rules for what each agent should deliver and how the results are passed on.
In the outer loop, the best workflow — the "champion" — is run 18 times per task. The model ranks the results itself and selects up to 15 runs as the basis for LoRA fine-tuning, meaning a lightweight update of the model weights. The model thus learns from its own best executions, ranked by its own taste.
The base model is Qwen3.6-27B, with Qwen Code as the harness. The task material consisted of twelve synthetic research tasks generated by GPT-5.5 in finance, robotics and pharmaceuticals — nine for training and three for testing. Of the nine training tasks, only eight yielded data, because the search found no workflow that beat the reference on one of the tasks.
The numbers — as reported by the researchers themselves
All performance figures come from the paper's authors; no independent evaluation exists among the available sources. The measurement "score per output token" therefore rewards not only raw quality but also efficiency — how good a score the model achieves per unit of generated text.
After two cycles, the following multipliers against the untouched base model are reported:
- MLR-Bench: 1.60x
- DSBench: 1.35x
- AstaBench: 1.28x
- ScienceAgentBench: 1.20x
On SWE-bench Verified and Terminal-Bench 2.0, both coding-heavy, the same system fell to 0.94x and 0.96x — that is, below the baseline. Head-to-head win rate against the base model rose from 53.9% after one round to 69.9% after two.
There is also a comparison point against a single-agent variant: According to AI Weekly, which cites the paper, the multi-agent setup reached a 68.3% win rate with 28.8 million supervised tokens, versus 42 million for the single-agent variant. "The MASS model thus achieves the higher score against this reference with 31% fewer supervised tokens," AI Weekly quotes.
It is worth noting that raw scores moved little in some places: The Decoder points out that DSBench went only from 0.473 to 0.482 — meaning the 1.35x multiplier primarily reflects an efficiency gain, that is, better score per token.
The researchers' own warning
Perhaps the most remarkable thing about the publication is what the authors themselves say about the method's weaknesses. Because the same model both produces and evaluates the results, a shared blind spot in a homogeneous loop can lead to flawed results being accepted as improvements and baked into the next model version.
"In a homogeneous loop, a shared blind spot can allow a flawed result to be produced, accepted as an improvement, and incorporated into the next model," the authors write, according to AI Weekly, which calls the risk "an Achilles' heel for recursive self-improvement." The Sakana formulation in the blog post points in the same direction: shared blind spots can allow errors to be accepted as progress and reinforced through training.
This is particularly relevant given the numbers: the four research benchmarks where MASS did best are precisely domains where the model itself can assess the quality of research work. On the coding benchmarks, where evaluation is more explicit and binary, the system fell below the baseline.
The interpretive dispute: self-improvement or calibrated taste?
There is disagreement about what the numbers actually show. The editor behind AI Weekly argues that the gains are in reality taste-calibration — that the model becomes better at producing what it can already evaluate well — rather than general self-improvement. The argument builds on the coding benchmarks falling below 1x, which suggests the improvement does not generalize across domains.
This interpretation is contested, and AIMag takes no position on who is right. But it points to a real question: if the gain primarily comes from the model learning to optimize against its own judgment, that is a different — and weaker — claim than the model generally becoming better at solving problems.
Caveats readers should know
Several factors limit how much can be concluded from this work:
- No independent evaluation. All figures are author-reported. The arXiv preprint (2610.12176) and code repository are referenced but not peer-reviewed, and AIMag has not run the code itself.
- Synthetic tasks. The twelve tasks were generated by GPT-5.5 in three sectors. How well this generalizes to real research problems is unverified.
- Limited scale. This concerns a 27B model, two cycles, and a relatively small task set — nine training tasks, of which eight yielded data. It is proof-of-concept, not a full-scale demo.
- Unclear first authorship. AI Weekly attributes first authorship to Hyunin Lee at UC Berkeley, while the Sakana blog lists Hyunin Lee, Jinglue Xu, Jeffrey Seely and Yujin Tang without stating an order, along with UC Berkeley colleagues Somayeh Sojoudi, Matei Zaharia and Donghyun Lee. The detail is unresolved, but not essential to the results.
What to watch next
Two things will determine whether MASS is a general principle or a domain-specific trick. First, whether successors replace the homogeneous evaluation with heterogeneous evaluators or external grounding — some form of ground truth that breaks the shared blind-spot risk the authors themselves warn about. Second, whether anyone outside Sakana AI and UC Berkeley reproduces the numbers, preferably on tasks that are not synthetically generated by another large model.