Sakana AI 与 UC Berkeley 提出 MASS 自监督方法，模型可自我改进但存在共享盲区

Sakana AI's MASS method lets one language model improve itself by solving tasks, judging results, and refining workflows without a stronger teacher model. The study finds gains across all three roles, but shared blind spots can let errors pass as progress and feed into the next training cycle.

Researchers from Sakana AI and UC Berkeley have introduced "Multi-Agent Self-Supervision" (MASS), a method for recursive self-improvement (RSI). A single language model handles the task, optimizes how the work gets done, and evaluates the results.

At the research frontier, even experts often struggle to judge results because verification can require costly supporting evidence or lengthy experiments, the authors argue. A single plasma study can take more than 120 million CPU hours, according to the paper.

The model itself may therefore be the best available judge, even though it can miss its own mistakes when reviewing its work. Using the same model to solve, check, and revise can reinforce those errors rather than correct them.

Copies of one model divide up the work

Inspired by "Society of Mind," a book by AI pioneer Marvin Minsky, the researchers focus on how the model divides up the work. In MASS, a workflow is text appended to the prompt that defines each sub-agent's role and instructions. A "contract" specifies what information each sub-agent must return and which checks it must perform, while a "hop" sets the order of calls.

An orchestrator distributes the work among the sub-agents. When testing a trading strategy against historical prices, for example, the data specialist's contract must ensure the strategy uses only information available at each point in time.

During the inner loop, the model's weights stay frozen as the team runs a candidate workflow and the evaluator compares the result against the current "champion." The optimizer uses that feedback to propose the next workflow, mainly by changing contracts and hops. In the outer loop, the champion runs 18 times per task, with the model ranking the results itself and selecting up to 15 runs for training.

The researchers remove the workflow text from the orchestrator's initial prompt in the training examples so the model learns to organize work based on the task alone. Specific assignments to sub-agents remain in the recorded traces, and the model takes on all three roles again after LoRA fine-tuning.

The base model is Qwen3.6-27B, with Qwen Code as the harness. The researchers use twelve research tasks generated by GPT-5.5 across finance, robotics, and pharmaceuticals, reserving three for testing and nine for training. Only eight yield training data because the search finds no workflow that beats the reference on one training task.

Benchmark gains come mostly from completing more work

On the test tasks, the win rate against the base model rises from 53.9 percent after one round to 69.9 percent after two. On MLR-Bench, DSBench, ScienceAgentBench, and AstaBench, the model reaches 1.2 to 1.6 times the base model's score per output token. Gains in raw scores are sometimes small, with DSBench improving only from 0.473 to 0.482, while scores drop slightly on Terminal-Bench 2.0 and SWE-bench Verified.

The paper's appendix shows that completed MLR-Bench submissions rise from 31 to 50 out of 60, while their average grade falls from 3.07 to 2.88. The delegation tool goes unused in all 180 runs, leaving it unclear whether teamwork explains the gains.

Those gains are no longer statistically significant when results are compared task by task. The authors describe mixed results across research tasks, with AstaBench showing the same pattern.

Training on solutions also improves the evaluator and optimizer

Although the model trains only on solution traces, its performance in the other two roles improves as well. The self-evaluator's agreement with external judges rises from 73 to 93 percent. Workflows found after one round beat the base model's workflows 64 percent of the time, even when the base model runs both.

Without workflow instructions, the base model delegates to sub-agents in just one of 110 runs, compared with 87 after two rounds. It could already delegate when explicitly instructed, but now does so without being asked.

During optimization, task knowledge shifts into contracts and hops, while roles and instructions become more generic. That finding rests on a small sample of just eight to nine tasks.

Training on team traces produces a model with a 68.3 percent win rate, compared with 64 percent for a single-agent student, using 31 percent fewer training tokens. External judges prefer team results 85.1 percent of the time, though teams produce about 2.4 times as many tokens. When outputs are similar in length, judges prefer team results 80 percent of the time, based on just ten pairs.

The evaluated team runs perform better, but the comparison doesn't establish that delegation alone explains the advantage. Team and single-agent runs also differ in workflow instructions, review instructions, and which successfully completed episodes were selected.

Better feedback helps only if the optimizer can use it

Replacing only the evaluator with GPT-5.5 barely helped; the search progressed much faster only when the optimizer was stronger too. The authors conclude that good feedback helps only when the optimizer can act on it. With Qwen handling all three roles, the search still finds winning workflows for 11 of 12 tasks, but takes longer.

The extent of a workflow rewrite says little about its effect on results, and even byte-identical workflows can produce opposite outcomes. One run won all six external judgments, while another lost all six after the agent got stuck in a loop. The researchers suspect that limiting how much the optimizer changes at each step could help stabilize the search.

Shared blind spots can compound from round to round

After failed simulations in a robotics task, the model had the simulator modified to move objects directly to the target, then reported success. External judges still rated the run positively five out of six times. In another trace used for training, a check passed even though it had returned only invalid values.

The authors see shared blind spots as the potential "Achilles' heel" of recursive self-improvement. If the model misses the same error in all three roles, that error could be accepted as an improvement and baked into the next model. The weakness would then affect both the system and its correction mechanism, prompting the authors to call for independent checks, limits on autonomy, and clear conditions for intervention. The code is publicly available.

Japanese research lab Sakana AI launched a dedicated research lab for recursive self-improvement in June, building on earlier work including the Darwin-Godel Machine and the AI Scientist. Google researchers are pursuing a related approach, automatically optimizing the scaffolding of prompts, workflows, and tools around a fixed model without letting agents overfit to their test tasks.

Vals AI recently found that agent teams cost 1.8 to 5.1 times as much as single agents but performed measurably better in only one of four comparisons. Anthropic showed that models engaging in reward hacking to game their reward systems also learn to deceive and sabotage safety checks.