一种融合图推理与多Agent协作的新思想生成路径识别框架
Framework of new idea emergent pathway based on graph reasoning and multi-Agent system
摘要
新思想如何产生是人类长期关心的话题，在传统科学发现的逻辑向AI4Science和Science4AI“双螺旋”驱动逻辑转变的背景下，如何利用AI Agent整合海量、异构的科学知识，发现并模拟人类新思想的产生过程，在以科学家经验和直觉为起点的科研发现向基于大数据和AI算法演进，让科学家从繁重的信息筛选转向高阶创造性思考，以及主动打破学科壁垒催生大量跨学科科学发现方面具有重要意义。本文提出一种融合图推理与多Agent协作的框架（Graph-reasoning And Multi-agent Pathfinding, GAMP）。该框架首先通过提示工程，从论文摘要中抽取三元组，并采用Neo4j进行存贮，形成大规模科学知识图谱作为结构化知识基底。然后，设计多个功能不同的 AI Agent（如领域专家Agent、路径探索Agent、创新性评估Agent等）组成协作系统，每个Agent由大模型驱动，赋予其语义理解、实体抽取、路径寻找等方面的能力。例如，在领域专家Agent中，通过知识库和提示工程，让该Agent专注于基因、蛋白、信号通路层面的合理性判断；在路径探索Agent中，采用广度优先算法、遗传算法、大模型引导的搜索等不同路径搜索方式，让该Agent专注于发现最新颖的路径。这些Agent在图上进行协同探索与推理，通过模拟科学团队的“头脑风暴”和“假设-验证”循环，生成并评估新思想的产生路径。我们以获得2021年诺贝尔生理学或医学奖的成果为例，收集1995年1月1日-2005年12月31日Web of Science核心合集数据库、Scopus以及Pubmed数据库中与温度触觉受体相关的文献作为案例，构建了“问题-方案-效果”三层知识网络进行实证研究。文章创新点在于：一是，连接符号主义与联结主义，充分发挥了图结构推理和大模型强大的语义理解的能力；二是，设计了结构化的多Agent协作协议，明确分工，模拟真实科研团队。局限性在于：形式化表示“新思想的路径”、深层次理解语义、新思想突破性潜力评估方面有待深入、细化。
Abstract
The question of how new ideas emerge has long captivated human curiosity. Against the backdrop of a paradigm shift from traditional scientific discovery logic to a "dual helix" driven by AI4Science and Science4AI, there is significant value in leveraging AI Agents to integrate massive, heterogeneous scientific knowledge and simulate the process of human ideation. This approach promotes the evolution of scientific discovery from reliance on scientist experience and intuition toward data- and algorithm-driven methodologies. It enables researchers to transition from laborious information filtering to higher-order creative thinking and actively breaks down disciplinary barriers, thereby fostering interdisciplinary discoveries.This paper proposes a framework integrating graph reasoning and multi-agent collaborationGraph-reasoning And Multi-agent Pathfinding (GAMP). The framework begins by using prompt engineering to extract triples from paper abstracts, which are stored in Neo4j to form a large-scale scientific knowledge graph as a structured knowledge base. Next, a collaborative system is designed comprising multiple functionally distinct AI Agentssuch as Domain Expert Agent, Path Exploration Agent, and Innovation Assessment Agenteach powered by a large language model (LLM) and equipped with capabilities including semantic understanding, entity extraction, and pathfinding. For example, the Domain Expert Agent, supported by knowledge bases and prompt engineering, focuses on evaluating plausibility at the level of genes, proteins, and signaling pathways. The Path Exploration Agent employs diverse search strategiessuch as breadth-first search, genetic algorithms, and LLM-guided searchto identify the most novel paths. These agents collaboratively explore and reason over the graph, simulating scientific team "brainstorming" and "hypothesis-validation" cycles to generate and evaluate paths for new idea formation.As a case study, we used the research recognized by the 2021 Nobel Prize in Physiology or Medicine, collecting literature on temperature and tactile receptors from the Web of Science Core Collection, Scopus, and PubMed databases between January 1, 1995, and December 31, 2005. A three-layer knowledge network"Problem-Solution-Effect"was constructed for empirical validation.The papers contributions are twofold: first, it bridges symbolism and connectionism, leveraging both graph-structured reasoning and the robust semantic comprehension of LLMs; second, it designs a structured multi-agent collaboration protocol with clear role specialization, simulating real-world research teamwork. Limitations include the need for further refinement in formalizing the "path of new ideas," achieving deeper semantic understanding, and improving the assessment of breakthrough potential for novel concepts.关键词
科学发现/图推理/大模型/AI AgentKey words
Scientific discovery/graph-reasoning/large language model/AI Agent引用本文复制引用
梁国强,林歌歌,张志豪,张硕.一种融合图推理与多Agent协作的新思想生成路径识别框架[EB/OL].(2025-10-13)[2026-04-03].https://chinaxiv.org/abs/202510.00069.学科分类
自然科学研究方法/信息科学、信息技术/计算技术、计算机技术
评论