Recursive self-improvement through collective intelligence (pub.sakana.ai)

Researchers have introduced Multi-Agent Self-Supervision (MASS), a groundbreaking framework for facilitating recursive self-improvement (RSI) in AI systems. MASS utilizes a single model to act as the executor, optimizer, and evaluator, creating a self-sustaining learning loop that enables AI to continuously learn from its own collective intelligence. The system first optimizes workflows based on self-evaluated performance, then trains on the best executions, leading to significant improvements in generating outputs, as measured by higher scores per output token across various benchmarks. The significance of MASS lies in its innovative approach to learning, akin to collective human efforts in knowledge advancement, where efficiency is gained through collaboration and division of tasks. By harnessing group intelligence within a multi-agent framework, MASS demonstrates substantial enhancements in task execution and collaborative learning, resulting in improved capabilities with fewer training tokens compared to traditional single-agent methods. This research sets the stage for AI systems capable of evolving beyond human expertise, representing a pivotal step toward realizing truly autonomous and intelligent agents in complex, non-verifiable domains.