Recursive Self-Improvement through Multi-Agent Self-Supervision

Sakana AI · 30 ● 3 sources

Multi-Agent Self-Supervision (MASS) trains a shared AI model using a team of virtual subagents that search for better workflows and then distill the model’s own selected results back into the shared model. After two cycles on synthetic open-ended research tasks with a 27B open-weights model, score per output token reached 1.2–1.6x the base model’s level across four research benchmarks. The approach changes training by adding recursive, self-evaluated multi-agent workflow selection, while raising concerns about reinforcing blind spots when the model evaluates its own outputs.