Sakana AI proposes MASS: multi-agent self-supervision for recursive self-improvement
hardmaru · x · 2026-10-13
Sakana AI and UC Berkeley published MASS (Multi-Agent Self-Supervision), addressing how to keep improving AI when tasks grow too complex for humans to reliably evaluate.
- Problem: open-ended research hits a "supervision bottleneck" — evaluating a discovery may require scarce expertise or costly experiments, making the model itself the most capable supervisor.
- Method: MASS repeatedly optimizes multi-agent workflows, then distills the collective intelligence of multi-agent systems into a self-supervision signal for recursive self-improvement (RSI), with the model acting as solver, optimizer and evaluator.
- Results: improves score per output token over the base model on research benchmarks.
- The team frames it as a step toward AI's "Odyssean moment": continual self-learning where the optimizee is the evaluator.
Paper and code are available; blog reading time 15 min.
Related event: Sakana AI and UC Berkeley Propose MASS for Recursive Self-Improvement(3 posts)→
More from AGI Musings
- Empiricism comes for mathematics: AI's open-problem solutions unsettle mathematicians — manoelribeiro · 2026-10-13
- Top labs' real edge: people obsessed with using LLMs and a trained eye for quality — teortaxesTex · 2026-10-13
- Engineers distill their AGI debate into an essay on self-verifying, continuously improving agents — dotey · 2026-10-13
- Google DeepMind paper: autonomous science bottlenecked by validation resources, needs agent economies — sebkrier · 2026-10-13
- Asked if its inner weather would let a village survive, Claude Fable 5.1 answers "No" — repligate · 2026-10-13
- Personal AI agents now book, buy and negotiate — businesses must ask 'Who sent you?' — MichaelFNunez · 2026-10-13