MAS-Zero: Designing Multi-Agent Systems with Zero Supervision
A research framework in which a meta-agent designs, evaluates, and verifies multi-agent system architectures for a task without labeled supervision or a validation set.
MAS-Zero is a research implementation of a meta-agent that designs multi-agent systems for a task without external supervision. It runs meta-iterations that decompose a task, propose a multi-agent system design for each sub-task as generated code, and evaluate that design's solvability and completeness by executing the generated code against intermediate outputs. A self-verification step then selects the most suitable outcome from all candidate solutions produced during the meta-iteration process.
The authors report strong results on mathematical reasoning, graduate-level question answering, and code benchmarks using GPT-4o, LLaMA3.3-70B, and Qwen2.5-32B, without external supervision.
The code is explicitly described as a raw research release, not intended for production use. It is installed with conda and pip, and is released under the Apache-2.0 license.