MASS: Recursive Self-Improvement through Multi-Agent Self-Supervision

gist

Sakana AI and UC Berkeley researchers present MASS, a recursive self-improvement loop in which one model serves as task solver, workflow optimizer, and evaluator: it searches for the best multi-agent workflow by its own judgment, then fine-tunes on the self-selected team’s trajectories. Two rounds on Qwen3.6-27B raise win rates on held-out research tasks to 69.9% and score per output token 1.2 to 1.6 times on research benchmarks, while the authors flag shared blind spots as a safety risk.

ideas

• Homogeneous RSI. On non-verifiable research tasks, the model being improved is also its own supervisor; GPT-5.5 and Claude Opus 4.8 judge progress only from outside the loop.
• Two loops. An inner loop evolves workflows of roles, instructions, contracts, and call order against a champion; an outer loop runs the champion 18 times, ranks workspaces with Bradley–Terry, and LoRA-tunes on the best trajectories without the workflow text.
• Role generalization. Training only on task-solving trajectories also improved the optimizer and evaluator, whose agreement with external judges rose from 73% to 93%; gains did not transfer to Terminal-Bench or SWE-bench.
• Information flows through contracts. Optimized workflows encode more task information in contracts and hops than in roles, and multi-agent trajectories gave a denser training signal, beating single-agent data with 31% fewer supervised tokens.
• The optimizer is the bottleneck. Swapping in a stronger evaluator helped a little, a stronger optimizer much more; the authors call for oversight that survives without a stronger supervisor.

quotes