Bridging Concepts: Multi-Agent Reasoning over Scientific Concept Graphs for High-Impact Research Prediction

Donggyu Yoon; Minsung Hwang; Juneha Baek; Donghyuk Shin; Joyce Jiyoung Whang

cs.AIOct 18, 2026

Bridging Concepts: Multi-Agent Reasoning over Scientific Concept Graphs for High-Impact Research Prediction

Authors: Donggyu Yoon, Minsung Hwang, Juneha Baek, Donghyuk Shin, Joyce Jiyoung Whang

Organizations: School of Computing, KAIST · College of Business, KAIST · Department of AI Computing, KAIST

Abstract

Within a research community, new topics often emerge when previously unconnected concepts are first studied together. We construct scientific concept graphs (SCGs)--domain knowledge graphs whose nodes are fine-grained research concepts extracted from paper text by LLMs and whose edges capture their co-occurrence in papers--and study the task of forecasting high-impact concept pairs, i.e., discovering concept pairs that are currently unconnected but will have high impact when connected. To this end, we introduce new benchmark datasets constructed from CIKM, KDD, and SIGIR conference papers, with varied prediction settings, including impact forecasting for emerging concepts unseen during training. We propose MiroGraph, a Multi-agent reasoning framework with scientific concept Graphs, in which a feature designer agent designs predictive features for a given concept pair from structural, semantic, and trend perspectives, and then three LLM agents compute the actual feature values from the enclosing subgraph of the concept pair. Using these feature values, a classifier designer agent constructs a classifier that predicts the final impact labels. MiroGraph outperforms various impact forecasting and temporal link prediction methods.

Figures & tables

Figure 1. Overview of MiroGraph. MiroGraph extracts an enclosing subgraph and employs a feature designer agent to determine predictive feature types based on statistics of the SCG. Three LLM agents compute feature values from structural, semantic, and trend perspectives. A classifier designer agent constructs classifiers to predict the final impact label based on the feature values. The figure illustrates the overall pipeline of MiroGraph. On the left, an enclosing subgraph is extracted for a candidate concept pair from a scientific concept graph. A feature designer agent then receives summary statistics of the entire SCG and determines structural, semantic, and trend feature types for the target concept pair. In the middle, three LLM-based agents, a Structural Analyzer, a Semantic Analyzer, and a Trend Analyzer, generate corresponding feature values from the enclosing subgraph. The generated features are normalized and combined into a unified feature representation. On the right, a classifier designer generates and selects classifiers using the generated feature representations from the training and validation sets to predict whether the concept pair is high-impact.

Venue	Setting	$\|\mathcal{V}\|$	$\|\mathcal{E}\|$	$\|\mathcal{P}\|$	$\|\mathcal{D}_{\text{train}}\|$	$\|\mathcal{D}_{\text{test}}\|$
SCG-CIKM	Test	6,544	74,887	2,254	1,916	338
SCG-CIKM	Eval	6,059	61,862	1,948	1,610	338
SCG-KDD	Test	3,833	41,685	3,196	2,716	480
SCG-KDD	Eval	3,529	34,561	2,490	2,010	480
SCG-SIGIR	Test	3,188	34,445	2,734	2,324	410
SCG-SIGIR	Eval	2,925	27,465	2,618	2,208	410

Table 1. Statistics of our SCG benchmark datasets.

|\mathcal{P}|

denotes the number of papers used to construct each snapshot graph.

Method	SCG-CIKM		SCG-KDD		SCG-SIGIR
	Test	Eval	Test	Eval	Test	Eval
Impact4Cast ( Gu and Krenn, 2024 )	0.6633 $\pm$ 0.0197	0.5865 $\pm$ 0.0106	0.6030 $\pm$ 0.0065	0.5811 $\pm$ 0.0074	0.6143 $\pm$ 0.0198	0.6106 $\pm$ 0.0063
TGAT ( Xu et al., 2020 )	0.6439 $\pm$ 0.0062	0.5465 $\pm$ 0.0069	0.5645 $\pm$ 0.0025	0.5662 $\pm$ 0.0013	0.5627 $\pm$ 0.0061	0.5527 $\pm$ 0.0148
DyGFormer ( Yu et al., 2023 )	0.6785 $\pm$ 0.0173	0.6264 $\pm$ 0.0178	0.5764 $\pm$ 0.0064	0.5736 $\pm$ 0.0081	0.6065 $\pm$ 0.0019	0.5951 $\pm$ 0.0076
CAWN ( Wang et al., 2021 )	0.6311 $\pm$ 0.0086	0.5699 $\pm$ 0.0049	0.5642 $\pm$ 0.0024	0.5727 $\pm$ 0.0018	0.5747 $\pm$ 0.0180	0.6001 $\pm$ 0.0098
GraphMixer ( Cong et al., 2023 )	0.5844 $\pm$ 0.0938	0.5162 $\pm$ 0.0227	0.5724 $\pm$ 0.0073	0.5194 $\pm$ 0.0502	0.5566 $\pm$ 0.0582	0.6118 $\pm$ 0.0131

Table 2. Main results in Macro F1 (

\uparrow

). Best results are bolded, and second-best results are underlined. An asterisk (*) indicates a statistically significant improvement over the best baseline based on a paired approximate randomization test (

p<0.05

Perspective	Feature	Value	Interpretation
Structural	Shortest path length	$4$	The pair is connected only through distant multi-hop bridges.
Structural	Common neighbors	$0.00$	The pair is not a trivial local-closure case.
Semantic	Novelty synergy	$0.75$	The combination suggests a novel cross-area integration.
Semantic	Cross-domain fit	$0.70$	The pair bridges educational modeling and graph representation learning.
Trend	Current relevance	$0.75$	Both concepts are relevant to recent research directions.
Trend	Attention durability	$0.73$	The topic is expected to sustain attention beyond a short-lived spike.

Table 3. Representative features and corresponding values for the pair Knowledge Tracing Methods and Heterogeneous Graph

Figure 2. Enclosing subgraph extracted for ‘ Knowledge Tracing Methods ’ and ‘ Heterogeneous Graph ’. For clarity, only a subset of nodes, edges, and edge attributes is visualized. The figure shows an enclosing subgraph extracted for a concept pair Knowledge Tracing Methods' and Heterogeneous Graph.' The two target concepts are connected through several intermediate research concepts, including Self-attention mechanism, Attention mechanism, Embedding Method, Graph neural networks, and Cold start recommendation. Edges are annotated with citation counts aggregated for recent years, illustrating the temporal changes in interactions among related concepts.

Method	SCG-CIKM ( Test )	SCG-CIKM ( Eval )
MiroGraph	$\bm{0.7501\pm 0.0217}$	$\bm{0.6435\pm 0.0032}$
w/o structure	$0.5844\pm 0.0041$	$0.5805\pm 0.0133$
w/o semantic	$0.7297\pm 0.0089$	$0.6192\pm 0.0196$
w/o trend	$0.7373\pm 0.0093$	$0.6065\pm 0.0010$
single score	$0.6939\pm 0.0050$	$0.5355\pm 0.0041$
MLP	$0.6928\pm 0.0204$	$0.5731\pm 0.0195$

Table 4. Ablation study on SCG-CIKM . Best results are bolded.

Knowledge Graphs

Knowledge graphs represent entities and relationships as structured, queryable networks for integrating data and supporting reasoning.

See all 230 papers

Time-Aligned Evolving Concept Graphs for Scientific Relation Forecasting
Sep 16, 2026Fred Sun, Jingze Wang, Minkun Xu +1AI-Assisted Scientific Research Temporal Link Prediction
Center for Brain-Inspired Computing Research, Department of Precision Instrument, Tsinghua University, Beijing, China · Guangdong Institute of Intelligence Science and Technology, Zhuhai, China
SCoR: A Hierarchical Framework for Forecasting Relations Between Scientific Concepts
Aug 27, 2026Jingze Wang, Fred Sun, Shangqi GuoScientific Discovery Temporal Link Prediction
Center for Brain-Inspired Computing Research, Tsinghua University
MIRAI: Prediction and Generation of High-Impact Academic Research
Jun 3, 2026Alex Li, Joseph JacobsonAI-Assisted Scientific Research Research Idea Generation
MIT Media Lab Cambridge, MA

Sep 16, 2026J/K move · Enter open · S save

Time-Aligned Evolving Concept Graphs for Scientific Relation Forecasting

Fred Sun, Jingze Wang, Minkun Xu, Shangqi Guo

Center for Brain-Inspired Computing Research, Department of Precision Instrument, Tsinghua University, Beijing, China · Guangdong Institute of Intelligence Science and Technology, Zhuhai, China

AI-Assisted Scientific Research Temporal Link Prediction Dynamic Graph Learning

Open paper AskPDF arXiv

Forecasting scientific relations can guide discovery by identifying promising connections before they emerge. Existing approaches often model concept semantics and graph structure separately or summarize semantics over coarse historical snapshots, leaving semantic representations potentially misaligned with rapidly evolving graph evidence. We propose a time-aligned evolving concept graph framework that jointly models semantic and structural evolution. Its core idea is to treat dated papers as shared update events, reconstructing semantic and structural states from the same publication history through each prediction time. Pair-level fusion combines these states to forecast first co-occurrence, relation formation, and conditional relation type. Holding architecture and training fixed, refreshing context alongside graph updates improves mean relation AUPRC by 16.6% over frozen context. On a graph built from 187,848 papers with 270,687 concepts and 7.45 million co-occurrence links, the complete framework improves mean relation AUROC from 0.9290 for the strongest evaluated baseline to 0.9722, with mean population-weighted AUPRC 0.005778.

Abstract

Knowledge Graphs

Explore similar work

Time-Aligned Evolving Concept Graphs for Scientific Relation Forecasting

SCoR: A Hierarchical Framework for Forecasting Relations Between Scientific Concepts

MIRAI: Prediction and Generation of High-Impact Academic Research