Mimosa: An Evolutionary Multi-Agent System Framework for Scientific Research

Mimosa is an evolutionary multi-agent system framework for scientific research. It enables automated collaboration through dynamic tool discovery, workflow topology generation by a meta-orchestrator, and execution of subtasks by code-generating agents. It addresses the limitations of current autonomous scientific research systems that rely on fixed workflows and tool sets, achieving a 43.1% success rate on ScienceAgentBench. Its core design philosophy is autonomous evolution, with features like modularity, auditability, and openness.

Autonomous Scientific Research (ASR) is a cutting-edge direction of AI in the scientific research field, driven by LLMs and agent architectures. However, most ASR systems rely on predefined fixed workflows and tool sets, making them difficult to adapt to dynamic changes in scientific research (such as the emergence of new methods and task evolution). When facing new tasks or needing to integrate new tools, manual reconfiguration is required.

Mimosa was proposed to address the limitations of ASR. Its core idea is to enable the system to automatically synthesize multi-agent workflows based on task requirements and optimize them iteratively through feedback. Key principles include: modularity and scalability (tool-agnostic design), auditability (complete recording of execution traces and version history), and openness (fully open-source).

The Mimosa architecture consists of four components:

1. Dynamic Tool Discovery Layer: Based on the MCP protocol, it dynamically identifies new tools at runtime to enhance adaptability;
2. Meta-Orchestrator: Generates customized workflow topologies for tasks, defining agent roles, data flows, and task decomposition;
3. Code-Generating Agents: Execute subtasks, call tools, or generate code to operate scientific libraries and resources;
4. LLM Evaluator and Feedback Loop: Scores execution results and generates feedback to drive iterative optimization of workflows.

In the ScienceAgentBench benchmark test, Mimosa achieved a 43.1% success rate using the DeepSeek-V3.2 model, surpassing baselines of single agents (lacking collaboration) and static multi-agent systems (fixed workflows). Different models show significant differences in response to multi-agent decomposition and iterative learning, suggesting that system design needs to consider model characteristic matching.

Mimosa completely records execution traces (workflow structure, agent input/output, tool calls, code results, etc.) and archives workflow version history. It supports post-hoc review, error analysis, and scientific reproducibility, making it easy for domain experts to guide corrections at key stages.

Mimosa's modular and tool-agnostic design makes it versatile, capable of handling interdisciplinary scientific research tasks such as bioinformatics and material simulation. As an open-source platform, it provides open infrastructure for the ASR field, supporting the community in developing new tools and improving algorithms to accelerate the transition of autonomous scientific research from concept to application.