By organizing data, tools and procedures in reusable bricks that improve in use, MatOS exceeds Codex CLI and Claude Code with an equal basic model on several scientific test benches.
A materials science study often involves several days of bibliographic analysis, calculations and experiments. LLM-based agents already know how to query databases or control simulation codes, but struggle to stay on track over time and learn from their past tasks. MatOS, presented on arXiv, tackles these limitations.
Reusable bricks that evolve
The system stores data, knowledge, tools, skills, roles, and workflows in a graph of scientific "assets." Each task is executed as an explicit graph with scientific controls, allowing it to be restarted and tracked at every step. Finally, assets are reviewed based on execution logs and user feedback.
On the ResearchClawBench, MatOS is up 11.15% compared to the previous best agent with the same base model, and is ahead of Codex CLI by 2.13% on MatTools and 4.17% on SciAgentGym. In eight case studies, it has reduced more than 30 million structures to 97 solid electrolyte candidates, identified 89 variants of halides that are more conductive than their original structures, and contributed to two new methods that reduce errors by 3.1% and 13.6%. The code is published as open source.
Read the full article on the original source
Go to the source →