Sakana AI and Berkeley propose MASS: multi-agent self-supervision for recursive self-improvement
TheTuringPost · x · 2026-10-13
Sakana AI and UC Berkeley propose MASS (Multi-Agent Self-Supervision), a recursive self-improvement (RSI) method where copies of the same model form a team, breaking the homogeneous self-critique bottleneck.
Two alternating loops:
- Inner loop: the model proposes, runs, and self-evaluates multi-agent workflows, revising instructions, task assignments, and information routing (weights frozen).
- Outer loop: optimized workflows generate trajectories used to supervised-fine-tune the shared model on both subagent and orchestrator traces.
Results over two MASS cycles with Qwen3.6-27B: 1.2–1.6x better score per output token on four open-ended benchmarks; evaluator accuracy up from 73% to 93%; 68.3% vs 64.0% win rate with fewer training tokens (a student trained on multi-agent traces beats one trained on single-agent traces with 1.4x more tokens); slightly worse per-token scores on coding/terminal tasks.
The authors note only two improvement rounds were shown, not endless self-improvement — but the insight stands: a model can improve by learning to split work across its own copies.
Related event: Sakana AI and UC Berkeley Propose MASS for Recursive Self-Improvement(3 posts)→
More from AGI Musings
- Bill Gates says the risk of AI "getting out of control" is still years away — nordicinst · 2026-10-13
- Anthropic evals: small models months from self-replicating on IoT; 9B model ran on 8 of 9 hardware types — scaling01 · 2026-10-13
- Foresight Vision Weekend 2026 lineup: OpenAI's Zaremba, Neuralink, DeepMind and more — allisondman · 2026-10-13
- Empiricism Comes for Mathematics: AI Proofs Break a Centuries-Old Alignment — manoelribeiro · 2026-10-13
- Counterparty agents will be the only way to survive the coming deluge of AI automation — robleclerc · 2026-10-13
- Gelsinger: AI speeds chip design but chips still take quarters; the whole flow needs rethinking — ai · 2026-10-13