Multi-Agent Self-Supervision (MASS) enables recursive self-improvement (RSI) by using a single AI model as both the optimizee and its own optimizer and evaluator. This homogeneous approach addresses the challenge of advancing beyond human knowledge, where human experts may struggle to assess results on non-verifiable tasks. MASS repeatedly optimizes multi-agent workflows, distills collective intelligence from multi-agent systems, and uses it as a self-supervision signal.
The optimizer searches for the best team workflow using evaluator feedback, then trains the executor on self-judged best team executions. The updated model fills all three roles and repeats the process, improving score per output token over the base model on research benchmarks. The approach draws inspiration from human societies, where collective achievements serve as learning starting points, and from Thomas Kuhn's concept of shared examples enabling scientific problem-solving.
Source: Hacker News · Summarized by HeadlinesBriefing