Position: Science is Collaborative—LLM for Science Should Be Too

Terry Jingchen Zhang, Wenyuan Jiang, David Guzman Piedrahita, Yongjin Yang, Sirui Lu, Bernhard Schölkopf, Zhijing Jin

Keywords: Multi-Agent Systems, Agentic AI for Science

Abstract:

Modern scientific breakthroughs are increasingly driven by collaborative team effort where researchers combine diverse expertise to tackle interdisciplinary challenges.
In this position paper, \textbf{we argue that LLM for Science should mirror such cooperative dynamics through Multi-Agent Systems (MAS) instead of pursuing a single omniscient model for all scientific problems}.
Following the Canonical Workflow Framework for Research (CWFR), we identify how MAS could benefit each canonical research stage: enhanced reliability for knowledge synthesis by cross-validation, increased creativity for hypothesis formulation via diversifying perspectives, improved robustness for experimental execution through parallel execution and fault-tolerant backup agents, and more diverse opinion in result interpretation and evaluation.
We further outline key bottlenecks in the current reality of MAS4Science and future work to address these challenges, concluding with several concrete call to actions for reliably scaling MAS in science from passive tools to active research partners.

Submission Number: 5