The approach targets a "supervision bottleneck" that occurs when tasks become too complex for human experts to guide progress. The researchers noted that this process of recursive self improvement could present safety risks, as a model acting as its own evaluator might reinforce its own blind spots. They urged the AI community to investigate how to detect and intervene in such failures as models begin to improve beyond human expertise.
Wednesday, October 21, 2026
Story
Sakana AI's New MASS Method Boosts 27B Model Performance 1.6x
AI9 days ago
Virtual subagents are used in a new training framework to help AI models overcome the limits of human evaluation when tackling open ended research. Developed by Sakana AI in collaboration with UC Berkeley, the Multi Agent Self Supervision (MASS) method organizes copies of a single model into teams to solve tasks and optimize workflows. When tested on a 27B open weights model over two cycles, the framework increased scores per output token by 1.2-1.6x across four research benchmarks, including MLR Bench, ScienceAgentBench, and DSBench.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.