MASS first searches for a stronger multi-agent workflow using the model’s judgments, then trains on executions from the selected workflow and repeats the cycle. The source says the method improves score per output token over the base model on research benchmarks. Training uses non-verifiable research tasks, so the loop’s supervision comes from the model’s own evaluations; the authors separately use GPT-5.5 and Claude Opus 4.8 to assess completed work, withholding those judgments from MASS. The reported results therefore provide an external assessment of progress, not direct verification of the model’s self-supervision signal.
Raw SignalOct 11, 20261 min read1 source
Sakana AI and UC Berkeley released MASS , a recursive self-improvement loop using one shared model for every role.
In one minute
Sakana AI and UC Berkeley introduced Multi-Agent Self-Supervision (MASS), a recursive self-improvement method in which one shared language model serves as task executor, workflow optimizer, and evaluator.