Mimosa: an open-source AI framework for autonomous scientific research featured in Le Monde
- Research
- Health
Published on May 22, 2026
–
Updated on May 22, 2026
Dates
on the May 20, 2026
Mimosa, the open-source multi-agent AI framework co-authored by Louis-Félix Nothias and Fabien Gandon, has been highlighted in a feature published on 20 May 2026 in Le Monde's science section, which explores how AI agents are accelerating drug discovery.
A research team from CNRS, Université Côte d'Azur, Inria, and 3IA Côte d'Azur has developed Mimosa, an open-source multi-agent AI framework that automatically generates and iteratively refines scientific workflows for computational research. The work is highlighted in a feature published on 20 May 2026 in Le Monde's science section, which explores how AI agents are accelerating drug discovery.
Mimosa addresses a core limitation of existing AI-driven research systems: their reliance on fixed, predefined workflows that cannot adapt when experiments behave unexpectedly or new tools become available. Rather than executing a static pipeline, Mimosa generates task-specific agent configurations, evaluates their outputs, and refines its own workflow architecture through iterative learning. Evaluated on ScienceAgentBench (a benchmark of 102 data-driven discovery tasks spanning bioinformatics, computational chemistry, geographic information science, and cognitive neuroscience) Mimosa achieved a 43.1% success rate in iterative-learning mode, outperforming both single-agent and static multi-agent baselines.
The project was led by Louis-Félix Nothias (CNRS researcher, Université Côte d'Azur, HolobiomicsLab), Junior Professor associated with the 3IA, who conceptualised the study, developed the methodology, and supervised the project. Fabien Gandon, Inria Research Director and 3IA Chairholder, contributed to the manuscript and has highlighted the role of web standards and semantic technologies, such as ontologies and knowledge graphs, in making agentic systems like Mimosa more interoperable and auditable across research environments.
Both Mimosa and its companion tool-management platform Toolomics are released under the Apache License 2.0. The preprint is available on arXiv (30 March 2026).
→ Read the Le Monde article
→ Read Accelerating scientific discovery with Co-Scientist and A multi-agent system for automating scientific discovery in Nature
→ Mimosa on arXiv
→ HolobiomicsLab
Mimosa addresses a core limitation of existing AI-driven research systems: their reliance on fixed, predefined workflows that cannot adapt when experiments behave unexpectedly or new tools become available. Rather than executing a static pipeline, Mimosa generates task-specific agent configurations, evaluates their outputs, and refines its own workflow architecture through iterative learning. Evaluated on ScienceAgentBench (a benchmark of 102 data-driven discovery tasks spanning bioinformatics, computational chemistry, geographic information science, and cognitive neuroscience) Mimosa achieved a 43.1% success rate in iterative-learning mode, outperforming both single-agent and static multi-agent baselines.
The project was led by Louis-Félix Nothias (CNRS researcher, Université Côte d'Azur, HolobiomicsLab), Junior Professor associated with the 3IA, who conceptualised the study, developed the methodology, and supervised the project. Fabien Gandon, Inria Research Director and 3IA Chairholder, contributed to the manuscript and has highlighted the role of web standards and semantic technologies, such as ontologies and knowledge graphs, in making agentic systems like Mimosa more interoperable and auditable across research environments.
Both Mimosa and its companion tool-management platform Toolomics are released under the Apache License 2.0. The preprint is available on arXiv (30 March 2026).
→ Read the Le Monde article
→ Read Accelerating scientific discovery with Co-Scientist and A multi-agent system for automating scientific discovery in Nature
→ Mimosa on arXiv
→ HolobiomicsLab
