Year: 2026
Type: preprint
Abstract: Large language models are increasingly used as natural-language interfaces for chemistry and materials research, but their utility in scientific workflows depends on their ability to connect user requests with executable tools, structured data handling, and verifiable outputs. Here, we present GVIM 2.0, a multi-agent research system designed to support computational and informatics workflows in chemistry and materials science. GVIM 2.0 integrates language-model-based task planning with domain-specific Skills, packagebacked computational tools, sandboxed execution, and artifact generation. On a 400-question public benchmark suite spanning ChemBench, MaScQA, and ChemLLMBench, the complete GVIM 2.0 workflow answered 325/400 questions correctly, achieving an accuracy of 81.25%, compared with 285/400 questions and 71.25% accuracy for a same-base API-only baseline. We further evaluated GVIM 2.0 on task-native workflows covering molecular solubility prediction, steel yield-strength prediction, reaction-table extraction, MS/MS spectral retrieval, chemical entity extraction, retrospective BACE inhibitor prioritization, Matbench experimental band-gap modelling, and temporal external validation of BACE1 activity prediction. These workflows generated inspectable outputs, including prediction files, metrics, figures, and reports. The results show that GVIM 2.0 can organize reproducible data-to-result workflows for selected chemistry and materials informatics tasks, providing a practical framework for tool-augmented scientific research assistance.
Source: ChemRxiv
Author: Kangyong Ma
Institution: Zhejiang Normal University
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223182505