Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 12 papers for 2026-08-14
This paper studies oversmoothing in deep graph neural networks and proposes HouseGNN. Instead of updating node embeddings through repeated propagation, HouseGNN uses the aggregated neighborhood message to estimate a reflection direction and then updates embeddings by a Householder reflection, mitigating depth-induced information collapse.
This work presents EGRL (Edge generation-guided relation-aware learning) for RNA-protein interaction prediction. It addresses data sparsity and cold-start by generating edges and learning relation-aware representations, overcoming limitations of homogeneous graphs or predefined meta-paths used by many GNN-based RPIP methods.
We propose a unified multi-dimensional benchmark for complex graph reasoning in large language models. Using a five-stage semi-automatic framework, the benchmark aims to cover data complexity and enable unified evaluation across text-based and code-based reasoning, addressing limitations of manually constructed datasets.
We introduce a Difference-of-Convex Regularizer for graph learning when the graph Laplacian is given. By expressing the regularizer as a difference of convex functions, this approach sidesteps the dense and ill-conditioned Laplacian pseudoinverse and enables scalable differentiable programming-based learning.
To tackle input heterogeneity, this paper presents SIGIL, a framework that maps any attributed graph to a fixed-dimension representation space. It constructs a structural interaction graph in which nodes represent feature dimensions and weighted edges capture inter-dimension interactions, enabling fully-inductive learning.
BEST-KAG is a multimodal knowledge-driven framework for building engineering standards question answering. It models standard knowledge with a multimodal knowledge graph and uses a large language model to support multi-clause reasoning with clause-level evidence linkage.
We propose a framework that constructs Dynamic Master Logic Models as Knowledge Graphs for complex system diagnostics, using Retrieval-Augmented Generation and large language models. The approach automates the creation of DML models from system descriptions and represents them as KG-DML, enabling scalable, automated diagnostics.
We rethink graph counterfactual explanations by introducing discrete diffusion inversion to generate faithful, sufficient, and understandable counterfactuals. The method navigates the discrete, combinatorial search space of graphs to produce minimal structural modifications that flip a model's prediction.
HSTGFormer is a Hyper Spatial-Temporal Graph Transformer for 3D human pose estimation. It reformulates spatial-temporal reasoning as localized coupled graph aggregation over joint-time nodes, enabling unified and efficient reasoning, and achieves improved performance over staged spatial and temporal models.
On the MatBench band-gap benchmark, this work demonstrates an autonomous agent that builds crystal-graph models and optimizes them through an autonomous LLM research loop. Without external pretraining, the agent designs and tunes the most accurate model for band-gap prediction.
This paper questions whether reported gains from GraphRAG are truly significant by identifying unbiased evaluation pitfalls in current setups. It highlights two critical flaws and proposes an unbiased evaluation framework to better measure the real gains of GraphRAG.
This paper proposes a data management approach that merges relational databases with graph data to provide a knowledge model to generative tools. By replacing foreign keys with reference values and using record addresses, the system improves performance and enables efficient link-following for LLMs.