Sakana's MASS model trained on its own agent-team runs wins 69.9% against its base after two rounds

The weights stay frozen in the inner loop. The team runs a candidate workflow, and the evaluator compares the result with the current champion ^[3]. For the training examples, the researchers strip the workflow text out of the orchestrator's first prompt, so the model has to learn to organize the work from the task alone. After LoRA fine-tuning it takes on all three roles again ^[5].

The model grades itself because outside grading is expensive. The authors argue that at the research frontier, verification can need costly supporting evidence or lengthy experiments, and a single plasma study can take more than 120 million CPU hours, according to the paper ^[13]. The cost is that the judge shares the solver's blind spots, and using one model to solve, check and revise can reinforce its errors ^[17]. The Decoder reports that the internal evaluator sometimes accepts flawed steps or manipulated solutions as successes and bakes them into the next model ^[12].

In our view, the outer-loop filter is loose. With 18 runs per task and up to 15 selected, at most three runs per task are discarded ^[19]. A flawed run the ranker rates well survives unless it lands in the bottom three. One gate does say no: a training task was dropped because the search found no workflow that beat the reference ^[11]. Training on solution traces alone also improved the evaluator and optimizer, The Decoder reports ^[18]. The sources do not include the evaluator's agreement figure.

The step from round one to round two is 16.0 points ^[20]. In the paper's appendix, completed MLR-Bench submissions rise from 31 to 50 of 60 while the average grade falls from 3.07 to 2.88 ^[14]. That is 51.7% completed going to 83.3% ^[21], with a 0.19-point lower grade ^[22]. DSBench moves only from 0.473 to 0.482, and Terminal-Bench 2.0 and SWE-bench Verified dip slightly ^[7]. According to The Decoder, the gains are no longer statistically significant when results are compared task by task ^[16].

The delegation tool went unused in all 180 runs, so it is unclear whether teamwork explains the gains ^[15].

For the numbers to transfer, your work has to resemble twelve research tasks generated by GPT-5.5 across finance, robotics and pharmaceuticals, where finishing more of the pipeline pays ^[10]. You also need a similar stack: a 27B base model in the Qwen Code harness ^[8]. Runtimewire adds that uneven benchmark transfer and just two cycles leave long-run reliability and real-world costs unresolved ^[23]. We would read the result as evidence that the pipeline shifts behavior toward finishing work. It does not yet show that the shift survives ten rounds of a judge trained on its own output.