Sakana's self-improving MASS model hits ~70% win rate but can't see its own blind spots
The Decoder · rss · 2026-10-13
Sakana AI's MASS method lets a single language model improve itself by solving tasks, judging the results, and refining its own workflows. After two rounds of self-improvement, it reaches a win rate of nearly 70% against the base model.
But the work exposes a core limitation: when the model misses the same errors across all three roles — solver, judge, and workflow refiner — it mistakes them for progress and reinforces them during training. The model cannot catch its own systematic blind spots, so initial capability caps how far self-improvement can go.
More from Research
- RT-SAFE benchmark: real-time embodied agents crash 12.3× more, safe success drops to 0.7% — Lianhuiq · 2026-10-13
- EpochGo: local-first multi-agent AutoML on Ollama, $0 API costs and 100% offline — Grouchy-Excitement28 · 2026-10-13
- RT-SAFE: Agents Hit 94.1% Task Success but Only 0.7% Safe Success — Lianhuiq · 2026-10-13
- AI-Powered Paper Mills Fuel Academic Fraud Crisis, Scholars Blame Publication Metrics — asusarla · 2026-10-13
- Deep learning turns oncoprotein K-Ras(G12C) into a self-labeling fluorescent protein tag — KevinKaichuang · 2026-10-13
- Build your own Jev-style decision model in 15 minutes with open models — Arindam_1729 · 2026-10-13