MASS distills multi-agent workflows into a single self-improving model

Multi-Agent Self-Supervision, or MASS, is an attempt by Sakana AI and UC Berkeley researchers to let a model improve on open-ended research work when nothing better is available to grade it. On tasks like writing a research report there is no automatic answer key, and the authors argue that at the frontier even human experts may struggle to judge the results. That leaves what they call homogeneous recursive self-improvement: the model being trained also designs its own training process and judges its own output. The danger, as Sakana puts it in its announcement, is that a model in that loop reinforces its own blind spots. MASS tries to get a better signal by having the model work as a team and learn from the team's best results.

The team is organised by a workflow, a written plan added to the task prompt. It assigns each subagent a role and instructions, a contract saying what it must hand back, including any required checks, and the order in which subagents are called. In one evolved finance workflow described in the Sakana write-up, the data specialist must report whether any future information leaked into the training data before later agents run a backtest. MASS runs two loops. In the inner loop the model's weights stay fixed: the team runs a candidate workflow, the same model compares the result with the best workflow found so far, and, acting as optimizer, rewrites the plan using that feedback. In the outer loop the winning workflow is run 18 times per task, the model ranks the results, and up to 16 of the best runs are kept. The workflow text is removed from the orchestrator's prompt, so the model has to learn to organise the work from the task alone, and the model is fine-tuned on both the orchestrator's and the subagents' conversations. The updated model then takes all three roles and the cycle repeats.

The researchers ran two cycles starting from Qwen's Qwen3.6-27B on 12 synthetic research tasks in finance, robotics and pharmacy, keeping three out of training. To measure progress from outside the loop, they had GPT-5.5 and Claude Opus 4.8 judge the work, and withheld those judgments from MASS. On the held-out tasks, the trained model's win rate against the original rose from 53.9% after one round to 69.9% after two. On four public research benchmarks (MLR-Bench, DSBench, ScienceAgentBench and AstaBench), score per output token reached 1.2 to 1.6 times the base model's, according to the paper. That measure rewards reaching a score with less generated text. Measured by raw score alone, the paper reports smaller gains on some of them, for example 28.7% to 31.8% on ScienceAgentBench and 0.473 to 0.482 on DSBench. Transfer to software engineering did not happen: score per output token was 0.96 times the base on Terminal-Bench 2.0 and 0.94 times on SWE-bench Verified, which the authors suspect could be fixed by adding software tasks to the training data.

The authors draw three findings from the runs. Training only on task-solving runs also improved the model as optimizer and as judge. Its agreement with the outside judges rose from 73% to 93%, and the original model performing workflows found by the first-round model beat the same model performing its own workflows in 64% of comparisons. Team transcripts also proved to be more efficient training data: a model trained on multi-agent runs reached a 68.3% win rate against the base after 28.8 million training tokens, while one trained on single-agent runs reached 64.0% after 42 million. And when they swapped in GPT-5.5 as the judge alone, improvement sped up only a little, while also making it the workflow optimizer helped a lot. The authors read this as a sign that the optimizer, not the evaluator, can limit how fast such a loop improves. They close by flagging the same risk they started with: when a model judges itself, a shared blind spot can let a flaw be accepted as progress and trained into the next version.