Mimosa Framework: Toward Evolving Multi-Agent Systems for Scientific Research
Open-source AI system automatically designs and refines its own workflows for scientific discovery.
A research team led by Martin Legrand and Louis-Félix Nothias has introduced the Mimosa Framework, an open-source platform designed to advance Autonomous Scientific Research (ASR). Unlike current systems constrained by fixed workflows, Mimosa automatically synthesizes and evolves task-specific multi-agent systems. Its core innovation is a four-stage process: a meta-orchestrator first designs a workflow topology; code-generating agents then execute subtasks by dynamically discovering and invoking tools via the Model Context Protocol (MCP); an LLM-based judge scores the execution; and this experimental feedback drives iterative refinement of the entire workflow.
On the ScienceAgentBench, Mimosa achieved a 43.1% success rate using the DeepSeek-V3.2 model, surpassing both single-agent baselines and static multi-agent configurations. The results indicate that the benefits of this evolutionary approach depend heavily on the underlying LLM's capabilities. Beyond benchmarks, Mimosa's modular, tool-agnostic architecture and its commitment to full auditability—preserving every step in logged execution traces—make it a practical foundation for community-driven scientific automation. Released as fully open-source, the framework aims to enable researchers to automate a broad range of computationally accessible tasks across disciplines, from chemistry to data analysis.
- Achieves 43.1% success rate on ScienceAgentBench using DeepSeek-V3.2, outperforming static multi-agent systems.
- Dynamically synthesizes workflows using a meta-orchestrator and refines them via an LLM judge's feedback.
- Fully open-source with auditable execution traces, built on the Model Context Protocol (MCP) for tool discovery.
Why It Matters
Provides a self-improving, auditable AI system that can automate complex scientific research workflows, accelerating discovery.