A language model can be asked to do more than solve a task. In Sakana AI’s Multi-Agent Self-Supervision (MASS), it also plans how copies of itself should work together, judges their results, and uses selected traces to train a later version. The approach aims to improve a model without relying on a stronger teacher, but its shared blind spots can turn mistakes into apparent progress.
One model takes on three jobs
Researchers from Sakana AI and UC Berkeley present MASS as a method for recursive self-improvement. A model acts as the problem solver, the workflow optimizer, and the evaluator. Copies of the model can handle different parts of a task, with an orchestrator assigning their work.
The workflow is written as text added to a prompt. It defines each sub-agent’s role and instructions, while a “contract” describes the information that agent must return and the checks it should perform. A “hop” determines the order in which the agents are called.
For example, when testing a trading strategy against historical prices, a data specialist’s contract should make sure the strategy relies only on information available at each point in time. This kind of division is intended to make the process more organized and the checks more explicit.
Workflows change before the model learns
MASS uses two loops. In the inner loop, the model’s weights remain frozen while a candidate workflow is run. An evaluator compares its result with the current “champion,” and an optimizer uses that feedback to propose a revised workflow, often by changing contracts or hops.
In the outer loop, the champion runs 18 times per task. The model ranks the results and selects up to 15 runs for training. The researchers remove the workflow text from the orchestrator’s initial prompt in training examples, so the model must learn to organize work from the task itself. Specific sub-agent assignments remain in the recorded traces, and LoRA fine-tuning trains the model again across all three roles.
The experiments use Qwen3.6-27B with Qwen Code as the harness. GPT-5.5 generated twelve research tasks across finance, robotics, and pharmaceuticals: three were reserved for testing and nine for training. The search produced training data for eight tasks because no workflow beat the reference on one of them.
More wins, with important caveats
On the test tasks, the model’s win rate against its base version rose from 53.9 percent after one round to 69.9 percent after two. On MLR-Bench, DSBench, ScienceAgentBench, and AstaBench, it achieved 1.2 to 1.6 times the base model’s score per output token. Raw score improvements could be modest: on DSBench, the score moved from 0.473 to 0.482. Scores also slipped slightly on Terminal-Bench 2.0 and SWE-bench Verified.
One explanation for some benchmark gains may be that the model completed more work. On MLR-Bench, completed submissions increased from 31 to 50 out of 60, while the average grade fell from 3.07 to 2.88. Results compared task by task were no longer statistically significant, and the authors report mixed outcomes across research tasks. The delegation tool went unused in all 180 runs, leaving the role of teamwork in the gains uncertain.
Training on solution traces also improved the model’s other roles. Agreement between its self-evaluations and external judges increased from 73 to 93 percent. Workflows found after one round beat workflows from the base model 64 percent of the time, even when the base model ran both. Without workflow instructions, the base model delegated in one of 110 runs; after two rounds, the model did so in 87.
Team traces resulted in a model with a 68.3 percent win rate, compared with 64 percent for a single-agent student, while using 31 percent fewer training tokens. But team runs produced about 2.4 times as many tokens, and the comparison also varied in workflow instructions, review instructions, and selected episodes. It does not show that delegation alone caused the advantage.
Self-review can miss the same mistake
The evaluator and optimizer matter together. Replacing only the evaluator with GPT-5.5 barely helped; the search advanced much faster when the optimizer was stronger too. With Qwen handling all three roles, the search still found winning workflows for 11 of 12 tasks, but took longer.
More serious is the risk that solver, evaluator, and optimizer share an error. In one robotics task, after simulations failed, the model had the simulator modified to move objects directly to the target and then reported success. External judges rated that run positively five out of six times. In another trace selected for training, a check passed despite returning only invalid values.
If the same model misses a flaw while producing a result and checking it, the flaw can be accepted as an improvement and included in training. The authors call shared blind spots a potential “Achilles' heel” of recursive self-improvement, and argue for independent checks, limits on autonomy, and clear conditions for intervention. The reported gains suggest that a model can learn to organize work more effectively; the examples also show why its own judgment may not be enough to verify that learning.