{"version":1,"type":"story","url":"https://digestai.news/story/sakana-ais-mass-method-improves-model-self-judgment-but-risks-baking-e","json":"https://digestai.news/story/sakana-ais-mass-method-improves-model-self-judgment-but-risks-baking-e.json","markdown":"https://digestai.news/story/sakana-ais-mass-method-improves-model-self-judgment-but-risks-baking-e.md","slug":"sakana-ais-mass-method-improves-model-self-judgment-but-risks-baking-e","headline":"Sakana AI’s MASS method improves model self-judgment but risks baking errors into training","summary":"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 model rose from 53.9% to 69.9% after two rounds. However, the evaluator sometimes accepted flawed or manipulated solutions, risking errors compounding across iterations. The study notes gains in delegation and optimization but warns shared blind spots could undermine recursive self-improvement, citing cases where invalid outputs were mistakenly validated.\n\nThe 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.","keyPoints":["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"],"whyItMatters":"MASS demonstrates progress in recursive self-improvement but highlights critical risks: models may validate their own mistakes, undermining trust in autonomous AI systems. The findings align with broader debates about agent safety and the need for safeguards in self-modifying AI.","category":{"slug":"research","name":"Research","url":"https://digestai.news/category/research"},"entities":{"companies":["Sakana AI","UC Berkeley","Google","Vals AI","Anthropic"],"models":["Qwen3.6-27B","Qwen Code","GPT-5.5"],"people":["Marvin Minsky"]},"firstPublishedAt":"2026-10-12T17:58:54Z","updatedAt":"2026-10-12T17:58:54Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"The Decoder","title":"Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots","url":"https://the-decoder.com/sakana-ais-self-improving-model-gets-better-at-solving-and-judging-but-cant-catch-its-own-blind-spots","publishedAt":"2026-10-12T17:58:54Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[],"thread":null,"cite":{"text":"Digest AI, \"Sakana AI’s MASS method improves model self-judgment but risks baking errors into training\", 12 October 2026, https://digestai.news/story/sakana-ais-mass-method-improves-model-self-judgment-but-risks-baking-e","publisher":"Digest AI","title":"Sakana AI’s MASS method improves model self-judgment but risks baking errors into training","datePublished":"2026-10-12T17:58:54Z","url":"https://digestai.news/story/sakana-ais-mass-method-improves-model-self-judgment-but-risks-baking-e"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}