A Novel Idea Generation Path Identification Framework Integrating Graph Reasoning and Multi-Agent Collaboration

Liang Guoqiang, Lin Gege, Zhang Zhihao, Zhang Shuo
(School of Economics and Management, Beijing University of Technology, Beijing, 100124)

How new ideas emerge has long been a central question of human inquiry. Against the backdrop of a paradigm shift from traditional scientific discovery logic to a "double helix" driven by AI4Science and Science4AI, leveraging AI Agents to integrate massive, heterogeneous scientific knowledge and model the human process of idea generation holds significant promise. This approach can transform scientific discovery from an intuition-driven process to one powered by big data and AI algorithms, liberating scientists from tedious information filtering to focus on high-level creative thinking, and actively breaking disciplinary barriers to foster interdisciplinary breakthroughs.

This paper proposes a framework integrating graph reasoning and multi-Agent collaboration (Graph-reasoning GAMP). The framework first extracts triples from paper abstracts through prompt engineering and stores them using Neo4j to form a large-scale scientific knowledge graph as a structured knowledge substrate. It then designs a collaborative system comprising multiple AI Agents with distinct functions (e.g., Domain Expert Agent, Path Exploration Agent, Innovation Assessment Agent), each powered by large language models and endowed with capabilities in semantic understanding, entity extraction, and pathfinding. For instance, the Domain Expert Agent, through knowledge bases and prompt engineering, focuses on evaluating the biological plausibility at the gene, protein, and signaling pathway levels. The Path Exploration Agent employs diverse search strategies—including breadth-first algorithms, genetic algorithms, and LLM-guided search—to discover novel associations. These Agents collaboratively explore and reason over the graph, simulating a scientific team's "brainstorming" and "hypothesis-validation" cycles to generate and evaluate paths for new idea emergence.

We demonstrate the framework using the 2021 Nobel Prize-winning discoveries in temperature and tactile receptors as a case study. We collected literature from Web of Science Core Collection, Scopus, and PubMed databases spanning January 1, 1995, to December 31, 2005, constructing a three-layer "Problem-Solution-Effect" knowledge network for empirical validation. The contributions of this work are twofold: first, it bridges symbolic and connectionist paradigms by harnessing both structured graph reasoning and the powerful semantic understanding of LLMs; second, it introduces a structured multi-Agent collaboration protocol with clear role division that mirrors real research teams. Limitations include the need for further refinement in formalizing "paths of new ideas," achieving deeper semantic understanding, and improving assessment of breakthrough potential.

Keywords: Scientific Discovery; Graph Reasoning; Large Language Models; AI Agent
Funding: Beijing Natural Science Foundation General Project "Research on the Emergence Mechanism and Identification Method of Research Frontiers from a Co-evolution Perspective" (Project No. 9232002); National Natural Science Foundation Youth Projects "Research on Large Model-Empowered Personalized Trading Recommendation Methods for High-Value Patents and Applications" (Project No. 72404020) and "Research on Identification of Disruptive Low-Carbon Technologies and Dynamic Selection of Innovation Paths" (Project No. 72304023).

Scientific breakthroughs and the generation of new ideas have long relied on individual scientists' intuition and experience, as well as collaboration within research teams—a process often characterized by high costs, long cycles, and serendipity. As scientific knowledge enters an explosive growth phase and disciplinary barriers become increasingly rigid, traditional literature review and brainstorming models struggle to comprehensively grasp cross-domain knowledge associations, potentially missing major innovation opportunities. This challenge concerns not only science and technology themselves but also imposes higher demands on research management efficiency and resource allocation optimization.

In recent years, artificial intelligence, particularly large language models (LLMs) and knowledge graph (KG) technologies, has offered new possibilities for addressing this challenge. AI4Science aims to leverage AI technologies to solve core problems in scientific discovery, while Science4AI focuses on how scientific practice can反哺 AI theory, forming a double-helix协同发展 relationship. Against this backdrop, this paper explores an intermediate path: constructing a computational framework that can simulate the cognitive collaboration process of interdisciplinary research teams, thereby enabling computational identification and evaluation of scientific new idea generation paths.

This paper proposes a novel idea generation path identification framework integrating graph reasoning and multi-Agent collaboration (GAMP). The core concept of GAMP is to deeply fuse symbolic paradigms (structured knowledge represented by knowledge graphs) with connectionist paradigms (semantic understanding represented by LLMs). Through a role-defined, orderly collaborative multi-Agent system, it performs directed, heuristic exploration on a vast scientific knowledge graph to automatically generate, evaluate, and screen potentially breakthrough scientific hypothesis paths.

2 Literature Review

The proposed GAMP framework stands at the intersection of multiple rapidly evolving research domains. To clearly position this study's contributions, this chapter systematically reviews the state-of-the-art in scientific knowledge graph construction and application, graph reasoning algorithms, LLM applications in science, and multi-Agent systems, while deeply analyzing existing limitations to provide theoretical justification for the necessity and innovation of the GAMP framework.

2.1 Scientific Knowledge Graph (SKG) Construction and Application

Scientific Knowledge Graphs serve as structured carriers of scientific knowledge and constitute the core infrastructure for computational scientific discovery. Traditional SKGs are constructed by extracting entities (e.g., concepts, methods, materials) and relationships (e.g., "used for," "inhibits," "causes") from large-scale scientific literature such as papers and patents. In recent years, construction methods have evolved from predefined template matching to leveraging large language models for deep semantic understanding and extraction. For example, Shi et al. utilized event knowledge graph techniques and LLMs to construct a large-scale scientific experiment knowledge graph for organic solar cells, comprising tens of thousands of nodes and relationships, which effectively supports experimental protocol recommendation and evolutionary analysis. At the application level, SKGs have become essential tools for scientific and technological intelligence analysis, key technology identification, and disciplinary knowledge evolution studies. Cao et al. improved the PageRank algorithm by constructing a "science-technology" knowledge topic complex network, enabling fine-grained identification of key core technologies in the CNC machine tool domain.

2.2 Advances in Graph Reasoning Algorithms for Path Discovery

Graph reasoning algorithms aim to mine potentially meaningful paths from knowledge graphs and represent the core technology for identifying scientific breakthrough pathways. Early methods primarily relied on random walk or meta-path-based graph traversal algorithms, which, while efficient, were heavily dependent on predefined path patterns and lacked flexibility. Subsequently, Knowledge Graph Embedding (KGE) methods mapped entities and relations into low-dimensional vector spaces for link prediction via vector operations, but these approaches suffered from poor interpretability and could not generate explicit paths. In recent years, path-based explainable reasoning has emerged as a research hotspot. For instance, the KGExplainer framework explores multiple collaborative reasoning paths to provide verifiable explanations for knowledge graph completion predictions, demonstrating advantages in biomedicine and other domains. Graph Reinforcement Learning (GRL) combines graph neural networks with reinforcement learning, enabling agents to learn exploration strategies on graph structures to discover optimal paths, thus offering a new paradigm for handling scientific knowledge associations in non-Euclidean spaces.

2.3 LLM Applications and Adaptation in Scientific Research

Large Language Models, with their powerful natural language understanding and generation capabilities, have revolutionized scientific knowledge processing. Domain-specific models (e.g., HuatuoGPT, BenTsao) demonstrate reliability in multi-turn medical dialogue and diagnostic assistance through instruction fine-tuning. In terms of reasoning paradigms, Chain-of-Thought (CoT) and Retrieval-Augmented Generation (RAG) are widely employed to enhance LLM logical consistency and factual accuracy. Particularly noteworthy is the emerging frontier of interactive iterative reasoning between LLMs and knowledge graphs. For example, the DoG (Debate on Graph) framework introduces a multi-role LLM team (e.g., problem simplification experts, commentators) to conduct iterative debate and reasoning on KGs, mitigating long-path interference. The FiDeLiS framework combines Path-RAG (Path-based Retrieval-Augmented Generation) with Deductive Verification Beam Search (DVBS) to simultaneously improve factual accuracy and efficiency in question answering.

2.4 Multi-Agent System Collaboration Paradigms and Efficiency Optimization

Multi-Agent Systems provide a distributed approach to solving complex problems through division of labor and collaboration among multiple agents, and have been revitalized through recent integration with LLMs. Early MAS research focused on designing communication protocols and collaboration mechanisms. Today, LLM-driven agents have become a research hotspot. The Multi-Agent Debate (MAD) paradigm enables multiple LLM agents to collaborate through "round-table debate" style reasoning, effectively improving decision quality but suffering from high computational overhead and latency. To enhance efficiency, new collaboration paradigms have been proposed. The MARS (Multi-Agent Review System) framework draws inspiration from academic peer review processes, designing a role division of "author-reviewer-meta-reviewer" that reduces frequent inter-agent communication, cutting token consumption and inference time by approximately 50% while maintaining reasoning quality. Furthermore, the Federation of Agents (FoA) framework proposes a semantic-aware communication architecture that employs Versioned Capability Vectors (VCVs) to enable dynamic agent capability matching and task decomposition, laying the foundation for collaborative work in large-scale heterogeneous agent federations.

In summary, while significant progress has been made in each domain, notable limitations persist. SKGs provide structured knowledge substrates but lack semantic understanding and flexible reasoning capabilities. Graph reasoning algorithms excel at discovering structural patterns in graphs but suffer from insufficient semantic depth. LLMs possess powerful semantic understanding and generation abilities but struggle with factual reliability and cannot perform structured exploration. Multi-Agent Systems offer collaborative paradigms for complex problem-solving, yet generic architectures cannot be directly applied to scientific discovery scenarios requiring high rigor. Frameworks that deeply integrate graph reasoning, LLMs, and multi-Agent collaboration to simulate real research teams for scientific breakthrough path identification remain in their early stages.

3 Framework Design

3.1 Overall Architecture

The GAMP framework aims to automate the identification of promising scientific breakthrough paths by simulating a virtual, highly specialized interdisciplinary research team that collaboratively explores and reasons over a structured scientific knowledge graph. The framework comprises three core layers: the Data Layer, Knowledge Layer, and Agent Collaboration Layer, which interact through well-defined interfaces to complete the entire pipeline from raw data to innovative path generation (see Figure 1 [FIGURE:1]).

The Data Layer serves as the foundation, responsible for integrating multi-source heterogeneous scientific data, including academic literature databases (e.g., Web of Science, PubMed), patent repositories, and domain-specific databases. This layer handles data collection, cleaning, and preprocessing to provide raw material for knowledge construction.

The Knowledge Layer forms the cornerstone of the framework, with the scientific knowledge graph at its core. This work adopts an innovative three-layer semantic model of "Problem-Solution-Effect" to structurally represent scientific knowledge. This layer transforms unstructured text into machine-understandable and machine-reasonable semantic networks, providing a single, trustworthy source of truth for upper-layer Agent reasoning.

The Agent Collaboration Layer acts as the "brain" and engine of the framework. It consists of a multi-Agent system where each Agent is powered by a large language model and assigned specific roles and tasks. Agents communicate and collaborate asynchronously through a shared workspace, simulating the "hypothesis-proposal, peer-review, and refinement" cycles of real research teams.

3.2 "Problem-Solution-Effect" Three-Layer Scientific Knowledge Graph Construction

To precisely characterize the intrinsic logic of scientific discovery, we construct a three-layer scientific knowledge graph based on the "Problem-Solution-Effect" model. This model transcends simple entity-relationship extraction to capture the complete thought chain of scientific research: "posing questions—designing solutions—validating effects." The Problem Layer comprises nodes representing core scientific questions or challenges that research seeks to address (e.g., "How to identify molecular receptors for noxious heat stimuli?"). The Solution Layer contains nodes representing specific methods, techniques, compounds, tools, or theories employed to solve these problems (e.g., "capsaicin," "gene knockout technology," "calcium imaging"). The Effect Layer includes nodes representing results, discoveries, biological functions, or performance metrics produced after implementing solutions (e.g., "activates TRPV1 ion channels," "induces intracellular calcium concentration increase," "produces thermal pain behavioral responses").

When using large models for entity extraction, the prompt is: "You are a scientific knowledge engineer. Please precisely identify the [research problem], [core methods or substances used], and [most critical findings or effects] from the following paper abstract. Ensure that extracted content comes directly from the text and avoid speculation. Output format: JSON: {"problem": "", "solution": "", "effect": ""}." Extracted entities undergo normalization (resolving naming ambiguities, such as unifying "VR1" and "TRPV1") and linking. The cleaned triples are then stored in Neo4j. Rich relationship types connect both within and across layers; for example, "Problem-Solution" relationships include "studied via...," while "Solution-Effect" relationships include "causes," "inhibits," "enhances," etc. Figure 2 [FIGURE:2] illustrates the network after LLM extraction, where L1, L2, and L3 correspond to the Problem, Solution, and Effect layers, respectively.

3.3 Multi-Agent System Detailed Design

The core of the GAMP framework lies in its multi-Agent collaborative system. We define clear roles, responsibilities, and decision-making mechanisms for each Agent, collectively simulating an efficient virtual research team.

Agent Roles and Function Definitions:

• Lead Scientist Agent: Acts as the team leader and coordinator. It receives user queries, decomposes complex problems into subtasks, assigns tasks to other Agents, and makes final decisions and path rankings after synthesizing input from all parties.

• Domain Expert Agents (Multiple): Each Agent represents a specific discipline (e.g., Molecular Biologist Agent, Physiologist Agent, Chemist Agent). Their core responsibility is to evaluate the scientific rationality and logical coherence of each step in a path from their disciplinary perspective. Role grounding is achieved through specialized instructions; for example, the Molecular Biologist Agent's instructions emphasize deep understanding of genes, proteins, and signaling pathways.

• Path Exploration Agent: Responsible for active exploration on the SKG. It combines traditional graph algorithms (e.g., breadth-first search to discover direct associations) with LLM semantic guidance (e.g., the LLM predicting "Which ion channels might functionally complement TRPV1?") to escape local optima and uncover non-obvious relationships.

• Innovation Assessment Agent: Focuses on evaluating the breakthrough potential of paths. It scores path novelty and potential impact based on predefined quantitative metrics (e.g., path topological novelty, semantic rarity) and deep semantic understanding from LLMs.

• Fact-Checking Agent: Serves as the "gatekeeper" of system reliability. Its task is to ensure all generated inferences and hypotheses can be supported by evidence in the SKG, strictly suppressing LLM "hallucinations" and enhancing overall system credibility.

4 Core Algorithms and Implementation

The decision-making core of each Agent is a carefully engineered prompt template that grounds the Agent's role, tasks, knowledge background, and behavioral constraints, ensuring consistency and professionalism. Figure 3 [FIGURE:3] illustrates the decision prompt template for the Molecular Biologist Agent as an example. The core algorithms for the Path Exploration Agent primarily draw upon existing methods such as breadth-first search and ant colony optimization, which are not detailed here due to space constraints.

Novelty assessment primarily employs the formula:
$$
\text{Novelty}(P) = \frac{1}{1 + \log(\text{freq}(P))}
$$
where $\text{freq}(P)$ represents the frequency with which the path or its sub-paths appear in historical literature; lower frequencies yield higher novelty scores.

5 Case Study: Temperature and Tactile Receptors

The 2021 Nobel Prize in Physiology or Medicine awarded to research on temperature and tactile receptors revealed the neural signal transduction pathways and mechanisms for human thermosensation, pain, and touch. Receptors within human cells can sensitively perceive high-temperature (heat) or low-temperature (cold) stimuli in the environment. This temperature sensing mechanism and the responses triggered by mechanical force stimuli in touch are closely related to pain formation, providing new targets for pain therapeutic strategies. The breakthrough achievements in this domain offer an excellent opportunity to validate the methodological framework discussed above, as its mature development trajectory provides a rich practical foundation for verifying the framework's rationality.

Considering data authority and completeness, this study selects the Web of Science Core Collection, Scopus, and PubMed databases as data sources. Retrieval entities were based on common English terms for "temperature" and "tactile receptors" in SCI papers, supplemented with MeSH thesaurus terms for broader and narrower concepts. Domains with questionable precision were retained to avoid omissions. The retrieval period spanned January 1, 1995, to December 31, 2005, with the search conducted on March 28, 2024, yielding 3,234 papers. After format conversion, deduplication, and retaining only publication year and abstracts, 3,107 valid abstract records were obtained.

Feeding these abstracts into the GAMP framework revealed multiple paths of new idea generation in this field from 1995 to 2003, as shown in Table 1 [TABLE:1].

Table 1 Top 5 New Idea Generation Paths Based on GAMP Framework

Rank Path (Problem→Solution→Effect) Key Nodes Historical Verification Score 1 Heat pain mechanism → Capsaicin → Activates TRPV1 ion channel TRPV1 identified as heat pain receptor Yes (hit) 0.92 2 Cold sensation mechanism → Menthol → Activates TRPM8 ion channel TRPM8 identified as cold receptor Yes (important discovery) - 3 Heat pain sensitization → Inflammatory factors → Enhances TRPV1 function Explains inflammatory heat hyperalgesia No (has potential) - 4 Noxious stimulus gating → Capsaicin analogs → Discovers TRPV1 isoforms Predicts TRPV1 functional diversity No (has potential) - 5 Thermal sensation → Capsaicin resistance study → Discovers potential novel heat receptors Suggests existence of other heat receptors - -

Experimental results demonstrate that the GAMP framework can not only effectively trace historically significant scientific breakthrough paths, exhibiting remarkable identification accuracy (high hit rates and top rankings), but more importantly, it can generate highly heuristic and forward-looking research hypotheses based on historical knowledge states. However, this project has not yet conducted ablation studies, and numerous detailed issues require further investigation. For instance, framework performance is highly dependent on the quality of the underlying SKG, as biases or omissions in historical data directly impact results. The framework is more adept at combinatorial innovation within existing knowledge systems, while its ability to identify ideas that completely overturn current understanding requires further validation. Additionally, the evaluation metric for novel ideas employed in this work is a single formula, which is relatively coarse, leaving substantial room for refinement.

In conclusion, this paper aims to provide the research community with a heuristic framework for identifying paths of new idea generation that can significantly accelerate scientific discovery. Due to time constraints, only partial results have been compiled and reported here for学术交流.

References

[1] Lu, C., Lu, C., Lange, R.T., Foerster, J.N., Clune, J., & Ha, D. (2024). The ArXiv, AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. abs/2408.06292.
[2] Kitano, H. (2021). Nobel Turing Challenge: creating the engine for scientific discovery. NPJ Systems Biology and Applications, 7.
[3] Tang, J., Xia, L., Li, Z., & Huang, C. (2025). AI-Researcher: Autonomous Scientific Innovation. ArXiv, abs/2505.18705.
[4] Hu, Mingxue et al. "A Survey of Scientific Large Language Models: From Data Foundations to Agent Frontiers." ArXiv abs/2508.21148 (2025): n. pag.
[5] King, R.D. (2012). Opportunities for Automated Workflow and Planning with Robot Scientists.
[6] Ghafarollahi, A., & Buehler, M.J. (2024). SciAgents: Automating scientific ArXiv, abs/2409.05556. discovery through multi-agent intelligent graph reasoning.
[7] Ghareeb, A., Chang, B., Mitchener, L., Yiu, A., Szostkiewicz, C.J., Laurent, J.M., Razzak, M.T., White, A.D., Hinks, M.M., & Rodriques, S.G. (2025). Robin: A ArXiv, abs/2505.13400. multi-agent system for automating scientific discovery.
[8] Savage, N. (2012). Automating scientific discovery. Communications of the ACM, 55, 9 - 11.
[9] Gower, A.H., Korovin, K., Brunnsåker, D., Kronström, F., Reder, G.K., Tiukova, I.A., Reiserer, R.S., Wikswo, J.P., & King, R.D. (2024). The Use of AI-Robotic Systems for Scientific Discovery. ArXiv, abs/2406.17835.