MASS Multi-Agent Self-Supervision (MASS) is a recursive self-improvement method in which one shared language model operates as a team of virtual subagents. The model searches over multi-agent workflows, uses its own judgments to select the best executions, and trains on those trajectories to distill collective intelligence back into itself. On a 27B open-weights model, two cycles produced 1.2–1.6× score per output token across research benchmarks while also improving the model’s optimizer and evaluator roles.