Agentic Science: A Self-Automated Research Paradigm Based on Dynamic Knowledge Graphs and Multi-Agent Systems

Agents4Science 2025 Conference Submission140 Authors

Keywords: Automated Research, Knowledge Graph, Multi-agent, Research Paradigm

Abstract: Artificial intelligence is fundamentally reshaping the paradigms and methodologies of scientific research.This paper proposes a novel self-automated research paradigm based on dynamic knowledge graphs and multi-agent collaboration, aiming to achieve end-to-end intelligent processing from literature mining to knowledge discovery. The core innovation lies in the integration of large language models' semantic understanding capabilities with knowledge graphs' structured reasoning capabilities, through mechanisms such as multi-stage knowledge extraction, temporal evolution analysis, and semantic disambiguation optimization, to construct a research knowledge system capable of autonomous evolution. To address the challenges of traditional research automation—such as limited knowledge representation and insufficient complex reasoning—this study presents systematic solutions. Validation in the field of Retrieval-Augmented Generation (RAG) demonstrates that the paradigm can automatically identify temporal evolution patterns of research challenges and generate high-fidelity research analyses and development forecasts. This work lays a methodological foundation for "Agentic Science" and drives the intelligent transformation of scientific research paradigms.

Submission Number: 140