Researchers at Sakana AI and UC Berkeley have published a method they say lets a language model improve itself by organizing, judging and learning from its own multi-agent teams — an approach that extends recursive self-improvement into research tasks where no automatic answer key exists.
The paper, "MASS: Recursive Self-Improvement through Multi-Agent Self-Supervision," appeared in October 2026 on arXiv (2610.12176) alongside an interactive project page from the Tokyo-based lab. Its author list spans both institutions: Hyunin Lee of UC Berkeley, who completed the work during an internship at Sakana AI; Sakana researchers Jinglue Xu, Jeffrey Seely and Yujin Tang; and UC Berkeley's Donghyun Lee, Somayeh Sojoudi and Matei Zaharia. For more context on this story, see our ongoing latest AI developments.
The work drew immediate attention in the developer community over the weekend, where a Hacker News discussion of the project passed 100 points within hours.
When the Model Is the Only Supervisor Available
Most recent progress in self-improving AI has relied on verifiable rewards: math answers that can be checked, code that either passes tests or does not. The Berkeley-Sakana team starts from a different premise. At the research frontier, they argue, there is often no ground truth at hand. Evaluating a proposed discovery may require scarce expertise, costly equipment or experiments that take far longer than generating the proposal itself.
The result is a widening gap: the ability to produce new work is advancing faster than the ability to supervise it. When no stronger teacher is available, the authors write, the model being improved — what they call the "optimizee" — may be the most capable supervisor left. That motivates what they term homogeneous recursive self-improvement, in which the same model serves as its own optimizer and evaluator.
The open question is how a system can generate a useful training signal for itself in that loop. Their answer borrows from human history.
How MASS Works
MASS, short for Multi-Agent Self-Supervision, turns one shared model into a three-role system. The same weights act as the task-solving agent that produces a workspace of code, results and a research report; the workflow optimizer that revises how a team of subagents divides the work; and the evaluator that compares two workspaces and judges which better satisfies the task.
The inner loop searches over team workflows — written plans covering each subagent's role, its instructions, the "contracts" specifying what evidence agents must hand one another, and the sequence of calls, or hops, between them. A champion workflow is retained across iterations and must be beaten to be replaced. One evolved finance workflow, condensed in the paper, sequences data preparation, feature construction, modeling and backtesting before an audit, and its data contract forces the data specialist to check that a backtest could not have leaked future information into historical decisions.
The outer loop then converts the team's collective experience into training data. The system runs the champion workflow 18 times, self-ranks the resulting workspaces, keeps the best 16 executions and fine-tunes the shared model on them — with the workflow prompt itself removed afterwards, so the coordination skills are carried in the weights rather than the instructions. The updated model then resumes all three roles, and the process repeats.
Measurable Gains on Non-Verifiable Tasks
The team ran two complete rounds of MASS starting from Qwen3.6-27B in the qwen-code coding-agent environment. The training ground was a suite of 12 synthetic research tasks spanning finance, robotics and pharmacy, each requiring code, quantitative results and a written report. Nine tasks were designated for training and three were held out from post-training; task descriptions were generated by GPT-5.5 before the experiment.
On the three held-out test tasks, the model's externally judged win rate against its base version rose from 53.9% after one round to 69.9% after two. The second-round model also beat the first-round model in 60.8% of comparisons. To avoid grading with the same yardstick used during training, the researchers had GPT-5.5 and Claude Opus 4.8 assess the completed work, with those judgments withheld from MASS itself.
Efficiency improved alongside quality. After two rounds, score per output token reached 1.2 to 1.6 times the base model's level across four public benchmarks: MLR-Bench, DSBench, ScienceAgentBench and AstaBench.
A Safety Question, Not Just a Capability Demo
The authors are explicit that the setup doubles as a safety probe. If automated research systems can compound their own capabilities, conditions that produce runaway improvement need to be understood early — the paper notes that the prospect of an intelligence explosion "gives further reason to study these conditions early," framing its central question as how to preserve meaningful oversight when no stronger supervisor is available.
The publication lands amid intensifying debate on that exact risk. In late September, Geoffrey Hinton, Yoshua Bengio and other leading scientists issued a public warning about the possibility of a runaway intelligence explosion, and research into agentic systems that conduct scientific work has accelerated across the field.
Limitations to Keep in Mind
The results come with caveats the authors acknowledge. The synthetic suite is small, and all training signal is self-generated, meaning the correctness of the model's own judgments is unknown by design — external judges are used only for measurement, not training. One training task was excluded after workflow search failed to find any workflow that beat its reference, a reminder that the loop does not always produce gains. And the external evaluations themselves relied on frontier models, which are imperfect proxies for ground truth.
What Comes Next
For Sakana AI, MASS extends a research agenda focused on automating the scientific process itself, following the lab's earlier work on AI-driven experiment generation and agentic review systems for checking research claims. If the collective self-supervision recipe transfers beyond synthetic tasks, the bottleneck it targets — supervising AI output at the frontier — will only grow more pressing as automated research systems multiply.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
