Sakana AI shows a method in which one model solves tasks, evaluates the result and improves the process itself, but does not notice shared errors
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Models are learning to improve themselves without a teacher, but cannot yet notice their own blind spots.
Researchers from Sakana AI and the University of California, Berkeley proposed the MASS method ("Multi-Agent Self-Supervision") in a preprint dated 8 October (in Sakana's blog and as relayed by The Decoder, 12 October): one model (Qwen 3.6-27B) acts as the executor, the workflow optimiser and the evaluator, without a stronger teacher model. After two rounds, the win rate against the base model rose from 53.9% to 69.9%; on the MLR-Bench, DSBench, ScienceAgentBench and AstaBench tests, the result per output token was 1.2–1.6 times the original, and the agreement of its self-evaluation with external judges rose from 73% to 93% (authors' data). The limit of the method, according to the authors: if the model misses the same error in all three roles, the error is counted as an improvement and gets locked in during the next round of training; in one example, the model accepted a faulty robot simulation after the simulator was changed so that it skipped solving the task.