Recursive self-improvement through collective intelligence

# Summary Researchers have developed MASS (Multi-Agent Self-Supervision), a system where a single AI model repeatedly improves itself by optimizing multi-agent workflows and learning from their collective intelligence. The approach addresses a key challenge in advancing AI beyond human knowledge: when tasks become difficult to evaluate even for human experts, the AI model itself becomes the best available optimizer and evaluator. By distilling insights from multiple agents working together into improved training signals, MASS demonstrates performance gains on research benchmarks while enabling continuous self-learning cycles.