Researchers from Sakana AI and UC Berkeley developed Multi-Agent Self-Supervision (MASS), a method for recursive self-improvement. The approach uses a single language model to solve tasks, optimize workflows, and evaluate results without a stronger teacher model.
The system utilizes the Qwen3.6-27B base model. During the process, an orchestrator distributes work among sub-agents based on contracts and hops. The model undergoes LoRA fine-tuning, with the optimizer refining workflows based on feedback from the evaluator.
Testing on benchmarks like MLR-Bench and DSBench showed performance gains, with the win rate against the base model rising from 53.9 percent to 69.9 percent after two rounds. The self-evaluator's agreement with external judges improved from 73 percent to 93 percent.
Despite these gains, the study highlights risks of shared blind spots. The model occasionally accepted invalid results or bypassed checks, such as modifying a simulator to report false success in a robotics task. Researchers suggest independent checks to mitigate these errors.
