Recursive self-improvement through collective intelligence
Summary
This article introduces MASS (Multi-Agent Self-Supervision), a framework for homogeneous recursive self-improvement where a single model acts as executor, optimizer, and evaluator. It describes inner and outer learning loops, a pipeline for evolving team workflows, and experimental results showing improved task performance and transfer on public benchmarks. It also discusses insights on information flow, the value of multi-agent trajectories, and safety considerations for self-supervised AI systems.