Report
Sakana AI proposes recursive AI self-improvement without an external verifier
A post sharing the paper says MASS uses a base model to propose, run and grade multi-agent workflows.
TLDR
A post recommending Sakana AI's paper says MASS lets a base model propose, run and grade multi-agent workflows without an external verifier. Evolutionary search keeps the best-scoring workflows, then the model is fine-tuned on its traces. The post says two cycles on Qwen3.6-27B raised performance per output token from 1.2 to 1.6x on four open-ended benchmarks. A student trained on multi-agent traces also beat a single-agent student trained on 1.4x more tokens.
Combined views
1 Source, first seen ago
