Recursive Self-Improvement through Collective Intelligence

A single shared model can act as a team of virtual subagents: it generates several workflows, solves tasks through them, then distills the best team's work back into itself. Tests with a 27-billion-parameter open model lifted scores per token to 1.2-1.6 times the base level after two cycles. The team warns that a model grading itself can reinforce its own blind spots.

Summary written with AI by HowTechMoves, based on reporting by sakana.