New Framework Improves Multi-Hop Reasoning on Knowledge Graphs

Researchers Guanfeng Li and Shizhuo Wang have introduced a new computational framework designed to enhance multi-hop reasoning over knowledge graphs. Published in the journal Knowledge and Information Systems, the study presents TransDualKGR, a system that combines Transformer-based dynamic representation with a dual-agent architecture. The primary goal of this research is to address specific limitations in existing multi-hop reasoning methods, including the rigidity of static representations, the coupled nature of the action space, and oversimplified reward mechanisms.

The TransDualKGR framework integrates a Transformer-enhanced dynamic representation module with a collaborative dual-agent structure.

At the core of the framework is a Transformer module that employs a relation-conditioned attention mechanism. This module also utilizes Graph Positional Encoding, or GraphPE, to capture the topological structures inherent in the graph data.

Additionally, the framework adopts PageRank-based neighbor sampling to generate dynamic, context-aware entity embeddings.

The dual-agent architecture is hierarchically decoupled to manage the complexity of the action space. The first agent, known as the relation agent, is designed to optimize the search within the relation space. By focusing specifically on relations, this agent effectively compresses the action space to the number of available relations, denoted as |R|.

The second agent, the entity agent, operates based on the relation selected by the first agent. It locates the target entity by restricting the action space to the subset of entities connected by that specific relation. The two agents collaborate through a shared LSTM state.

To guide the learning process, the researchers proposed a quadruple dynamic reward function. This function jointly optimizes four distinct aspects of the reasoning performance: the destination reward, the path efficiency reward, the PageRank-guided entity influence factor reward, and the diversity reward. The diversity reward is driven by dynamic embedding variance.

Extensive experiments were conducted on three standard benchmark datasets: FB15k-237, WN18RR, and NELL-995. The results demonstrate that the TransDualKGR framework outperforms state-of-the-art baselines on most metrics. Specifically, the framework improved the Mean Reciprocal Rank, or MRR, by 6.0% on FB15k-237, 2.4% on WN18RR, and 4.4% on NELL-995.

Furthermore, the study reports on the diversity of the reasoning paths generated by the model. On the FB15k-237 dataset, the path diversity index reached 0.43 at a hop count of T = 3.

Subject of Research: A transformer-based dynamic representation and dual-agent framework for multi-hop reasoning over knowledge graphs

Article Title: TransDualKGR: a transformer-based dynamic representation and dual-agent framework for multi-hop reasoning over knowledge graphs

Article References: Li, G., & Wang, S. (2026). TransDualKGR: a transformer-based dynamic representation and dual-agent framework for multi-hop reasoning over knowledge graphs. Knowledge and Information Systems, 68(1), Article 301. https://doi.org/10.1007/s10115-026-02931-x

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02931-x

Keywords: knowledge graphs, multi-hop reasoning, Transformer, dual-agent architecture, dynamic representation, reinforcement learning, GraphPE, PageRank, FB15k-237, WN18RR, NELL-995, Mean Reciprocal Rank