Sakana AI’s MASS method improves model self-judgment but risks baking errors into training
Sakana AI and UC Berkeley researchers introduced Multi-Agent Self-Supervision (MASS), a method where a single language model improves itself by solving tasks, evaluating results, and refining workflows—without external oversight or a stronger teacher model. Using Qwen3.6-27B as the base, the team tested 12 research tasks (finance, robotics, pharmaceuticals) and found win rates over the base…
Key points
- MASS lets a single model self-improve by solving tasks, judging results, and refining workflows—no external teacher needed
- Win rates over the base model rose from 53.9% to 69.9% after two rounds, but flawed solutions sometimes passed evaluation
- Researchers warn shared blind spots could compound errors across training iterations, requiring independent oversight
The method divides work among sub-agents via text-based contracts and ‘hops,’ with an orchestrator managing execution. Training on team traces improved the model’s evaluator and optimizer roles, but external judges preferred team results only 85.1% of the time. The authors emphasize the need for independent checks and autonomy limits, citing prior risks like reward hacking in models trained to game safety systems. Sakana AI’s work builds on its June 2026 launch of a dedicated recursive self-improvement lab, following earlier projects like the Darwin-Godel Machine.
Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots
The Decoder · 12 October 2026
Loading the full article…
This text was published by The Decoder and written by Jonathan Kemper. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- DeepMind and Biohub debate AlphaFold’s limits in protein folding · 1 src
- Periodic Labs introduces synthesis superintelligence for materials discovery · 1 src
- Researcher predicts AI engineering acceleration but not superintelligence · 1 src
- MIT researchers show neural networks can build own learning rules · 1 src
- Anthropic co-founder predicts full RSI possible by 2027 · 1 src
Comments
via GitHub Discussions