Sakana AI reports research benchmark gains from self-trained agent teams

Sakana AI reports that its Multi-Agent Self-Supervision method improved score per output token to 1.2-1.6× the base model's level across four research benchmarks after two training rounds. The experiments started with Qwen3.6-27B and used research tasks in finance, robotics and pharmacy. The method uses one shared model to solve tasks, revise how agent teams work and judge their results. It then trains that model on selected team executions and repeats the process. Separate assessments by GPT-5.5 and Claude Opus 4.8 were withheld from the training loop. The gains did not extend across all workloads: score per output token was 0.96× the base value on Terminal-Bench 2.0 and 0.94× on SWE-bench Verified.