Sakana AI introduces Multi-Agent Self-Supervision (MASS) for homogeneous recursive self-improvement (RSI): one shared language model acts as task executor, workflow optimizer and evaluator. MASS searches multi-agent workflows, distills collective intelligence from team executions into training data, and repeats—aiming to improve score per output token on research benchmarks without verifiable automatic rewards.
Motivation
At the frontier, human experts may struggle to assess non-verifiable research outputs. MASS asks whether a model can learn from its own team’s collective intelligence—analogous to how scientific communities pass on shared examples (Kuhn) and how coordinated workshops produce knowledge beyond any single worker.
How MASS works
Inner loop (fixed weights): An optimizer revises multi-agent workflows (roles, instructions, contracts, hops) using evaluator feedback; the evaluator compares two workspaces and picks the better one.
Outer loop: The model is fine-tuned on executions from the self-judged best team; the updated model resumes all three roles.
Workflows are text plans added to prompts; evolved finance workflows, for example, encode leakage checks in output contracts.
Training uses non-verifiable research tasks; external GPT-5.5 and Claude Opus 4.8 judges measure progress but do not guide training.
Key results (from publication)
- M+ vs single-agent: 68.3% win rate after 28.8M supervised tokens vs 64.0% for single-agent S++ after 42M tokens (~31% fewer supervised tokens).
- External judges prefer multi-agent outputs in 85.1% of comparisons vs single-agent runs; team runs generate ~2.4× more tokens on average.
- Agreement between the internal evaluator and external judges rises from 73% to 93% across three cycles.
- Workflow-search ablations: upgrading both optimizer and evaluator to GPT-5.5 finds winning workflows for all 12 tasks within 5 iterations; homogeneous loop eventually reaches 11/12 but more slowly—suggesting the optimizer, not the evaluator, can be the RSI bottleneck.
- Optimal workflows shift task information from roles/instructions toward contracts and call order (mutual information analysis).
Safety note
The authors warn that shared blind spots in a homogeneous loop could let flawed outputs be accepted as improvements; they tie this to chain-of-thought monitorability and intelligence-explosion oversight research.
Published October 2026. † Hyunin Lee — work done during internship at Sakana AI; affiliations include UC Berkeley and Sakana AI.