Recursive Self-Improvement through Collective Intelligence

Sakana AI introduced Multi-Agent Self-Supervision (MASS), a method where a shared model uses a team of virtual subagents to solve tasks, searches for better multi-agent workflows, and selects the best ones using its own judgments.

Training on the best team's executions distills collective experience back into the shared model.

In two cycles using a 27B open-weights model on synthetic open-ended research tasks, the score per output token reached 1.2-1.6x the base model's level across four research benchmarks.