Skip to content
NewsResearch

Sakana AI and UC Berkeley train one model to execute, grade, and improve itself

· by Pondero Newsdesk

The short version

Sakana AI and UC Berkeley researchers published MASS in October 2026, a framework where a single model fills the executor, optimizer, and evaluator roles in a self-improvement loop, lifting win rate from 53.9% to 69.9% over two rounds.

Sakana AI and UC Berkeley train one model to execute, grade, and improve itself

A single model just played coach, player, and referee in its own training loop, and the referee's calls got measurably more trustworthy with each round.

What

Sakana AI, working with UC Berkeley researchers, published Multi-Agent Self-Supervision (MASS) in October 2026, a framework in which one model fills three roles at once: it executes a task as part of a multi-agent team, searches for a better way to organize that team, and judges which version of the team's work is best. The model then retrains itself on the winning team's conversation logs, and the updated model resumes all three jobs for the next round. Starting from Qwen3.6-27B inside Sakana's qwen-code agent environment, two rounds of this loop lifted win rate against the base model from 53.9% to 69.9% on held-out synthetic research tasks spanning finance, robotics, and pharmacy, per the paper. On four public research benchmarks (MLR-Bench, DSBench, ScienceAgentBench, and AstaBench), score per output token reached 1.2x to 1.6x the base model's level. The gains did not carry over to coding tasks: Terminal-Bench 2.0 and SWE-bench Verified scores slipped to 0.96x and 0.94x of the base model, per the same source.

The self-grading got more trustworthy, not just more skilled

The researchers also tracked whether the model's own judgments could be trusted as it rewrote itself, separately from whether its output improved. They checked the self-evaluator's calls against two outside models, GPT-5.5 and Claude Opus 4.8, whose assessments were withheld from training so they could not be gamed. Agreement between the self-evaluator and those outside judges climbed from 73% after the first round to 93% after the second, per the paper. Training on multi-agent conversation logs was also more efficient than training on single-agent logs: a model trained on team trajectories hit a 68.3% win rate using 28.8 million supervised tokens, while one trained on solo trajectories needed 42 million tokens to reach a lower 64.0% win rate, about 31% fewer tokens for a better result. For anyone building self-improving agent pipelines, that is the headline tradeoff: multi-agent training data appears to pack more usable signal per token than single-agent logs, even when the underlying model and tasks are identical.

The paper's own authors flag a limit worth tracking. A homogeneous loop where the same model both produces and grades its work risks a "shared blind spot," where a flawed result gets accepted as an improvement and baked into the next version. If humans also miss the flaw, later rounds can reinforce it rather than correct it.

What to watch next

Sakana opened a dedicated Tokyo research group for self-improving AI, called RSI Lab, in June 2026, alongside earlier projects like the Darwin Godel Machine and LLM-Squared, per The Decoder. Whether MASS becomes part of that lab's published roadmap, and whether a follow-up closes the software-engineering transfer gap by adding coding tasks to the training mix as the authors suggest, are the two open threads from this release.

Sources