MAS-Zero - A multi-agent system design framework from Salesforce

What is MAS-Zero?

MAS-Zero is a multi-agent system (MAS) design framework from Salesforce that automatically designs and optimizes MAS without human supervision. Based on meta-level design, the framework dynamically generates, evaluates, and improves MAS configurations for each problem instance during inference. The core of MAS-Zero lies in its meta-iterative process, which includes task decomposition, generation of corresponding MAS implementations, and meta-level feedback based on solvability and completeness. It selects the most suitable result from all candidate solutions based on self-verification. MAS-Zero performs exceptionally well in mathematical reasoning, graduate-level question answering, and code benchmarking, all without relying on any external supervision.

Main functions of MAS-Zero

• Automatic design of multi-agent systems (MAS)It eliminates the need for humans to manually design agent roles and communication protocols, automatically generating suitable MAS structures based on specific problems, thereby improving the system's adaptability and performance to new tasks.
• Dynamic adaptation problemDuring the inference phase, dynamic agent composition and problem decomposition are performed for each problem instance, enabling MAS to better cope with complex and ever-changing tasks.
• No validation set supervision requiredIt reduces the need for data and improves the flexibility and usability of the system by adjusting and optimizing the MAS configuration without relying on pre-labeled validation sets.
• Performance optimizationBased on meta-level design and self-verification mechanism, the MAS design is continuously iterated and improved to enhance the accuracy and efficiency of the system on complex tasks while maintaining cost-effectiveness.
• Self-evolution abilityDuring the reasoning process, jiyu provides self-feedback and evaluation, learns and evolves automatically, and gradually optimizes the design and performance of MAS without the need for external supervision signals.

MAS-Zero's technical principles

• Meta-Iterations:
  • Task decomposition and MAS generationThis approach decomposes complex problems into multiple subtasks and generates a corresponding MAS implementation (in executable code form) for each subtask. MAS design is treated as a code generation problem, and task decomposition and MAS configuration are dynamically adjusted based on a meta-agent.
  • Meta-level feedbackThe process involves evaluating the solvability and completeness of the generated MAS design, obtaining intermediate outputs by executing the MAS code, and analyzing the outputs to determine whether the current MAS can effectively solve the problem. Feedback is generated based on the evaluation results to guide subsequent iterative improvements.
• Self-VerificationAfter multiple iterations, the most suitable result is selected from all candidate solutions. By comparing the candidate answers generated in different iterations and combining certain verification strategies (such as majority voting, excluding invalid answers, etc.), the final output answer is determined.
• LLM-based meta-proxyBased on a large language model (LLM) as a meta-agent, it endows the system with the ability to understand and generate natural language instructions. During the meta-iteration process, the meta-agent is responsible for task decomposition, MAS generation, feedback generation, and final result verification. Its interaction with the LLM enables the optimization and improvement of the MAS design.
• Self-monitored learningThe entire process does not rely on external validation sets or human supervision; it performs self-supervised learning based on the system's own execution results and feedback signals. The meta-agent adjusts the MAS design based on the solvability and completeness of intermediate outputs, gradually improving the system's performance and adaptability.

MAS-Zero project address

• Project official website:https://mas-design.github.io/
• GitHub repository:https://github.com/SalesforceAIResearch/MAS-Zero
• arXiv technical paper:https://arxiv.org/pdf/2505.14996

Application scenarios of MAS-Zero

• Solving complex problemsIn fields such as mathematics and scientific computing, complex problems are broken down into multiple sub-tasks to improve the efficiency and accuracy of solving them.
• Natural Language ProcessingUsed in advanced question-answering systems and text generation and editing to produce high-quality natural language processing results.
• Software EngineeringIn code generation, optimization, and software testing, tasks are broken down into multiple subtasks to improve the efficiency and quality of software development and testing.
• HealthcareIt is used in disease diagnosis and treatment planning to improve the accuracy and personalization of medical decisions.
• EducationIt can be used for personalized learning and intelligent tutoring, breaking down learning tasks according to students' needs, and having different intelligent agents provide customized learning paths and tutoring content to improve learning outcomes.