← Home

Daily arXiv Papers

Graph Neural Networks · Graph Learning · LLM × Graph

Showing 12 papers for 2026-06-08

Spatiotemporal Imputation with Graph-Informed Flow Matching
GNN Graph Learning

Spatiotemporal data often have missing values, and traditional RNN/GNN approaches propagate errors across time and space. The paper GiFlow introduces a diffusion-based imputation method that uses graph-informed flow matching to better leverage spatiotemporal structure, reducing error accumulation and improving efficiency and accuracy.

Graph Neural Network leveraging Higher-order Class Label Connectivity for Heterophilous Graphs
GNN Graph Learning

Graph neural networks excel on homophilous graphs but falter on heterophilous ones, where connected nodes often have different labels. This work proposes a GNN that explicitly leverages higher-order connectivity among class labels to capture complex label relationships in heterophilous graphs, improving node classification performance.

Learnable Kernel Density Estimation for Graphs and Its Application to Graph-Level Anomaly Detection
Graph Learning

LGKDE presents a learnable kernel density estimation framework for graphs, where a graph neural network represents each graph as a discrete distribution over learned kernels. This moves beyond handcrafted kernels, enabling a unified treatment of structural patterns and semantic variation, and is applied to graph-level anomaly detection.

E2Former-V2: On-the-Fly Equivariant Attention with Linear Activation Memory
GNN Graph Learning

EGNNs face scalability bottlenecks from explicit geometric feature construction on every edge. E2Former-V2 offers a scalable solution by integrating algebraic sparsity with hardware-aware execution, introducing Equivariant Axis-Aligned Sparsification (EAAS) to prune computations while preserving equivariance.

Learning to Execute Graph Algorithms Exactly with Graph Neural Networks
GNN Graph Learning

The paper proves exact learnability results for executing graph algorithms with GNNs under bounded-degree and finite-precision constraints. It uses a two-step process: training an ensemble of MLPs to execute local node instructions, and then using this ensemble as the update function inside a GNN during inference to perform the algorithm exactly.

The Post-GCN Decade Revisited: Curvature-Stratified Evaluation of Relational Learning
Graph Learning Graph Theory

Current relational-learning evaluations rely on flat leaderboards that overlook geometry. The paper argues that intrinsic geometry is a key latent factor influencing model effectiveness, and introduces curvature-stratified evaluation to reveal geometry-dependent performance variations and prevent misleading conclusions from aggregated metrics.

ADAGE: Active Defenses Against GNN Extraction
GNN Graph Learning

ADAGE proposes active defenses against GNN extraction attacks, addressing the risk of model stealing. The framework monitors signals from potential extraction attempts and deploys defense mechanisms to degrade attacker success while preserving legitimate model utility.

Understanding Generative Recommendation with Semantic IDs from a Model-scaling View
LLM × Graph

Generative Recommendation with Semantic IDs (SIDs) uses discrete codes from modality encoders to represent items in autoregressive user interaction modeling. This work analyzes how model scale affects SID-based GR and examines how SIDs encode item semantics and collaborative signals for recommendation.

Towards Iterative End-to-End Software Development: A Feature-Driven Multi-Agent Framework
Graph Learning

EvoDev is an iterative feature-driven multi-agent framework for end-to-end software development. It decomposes requirements into user-valued features and coordinates multiple agents to implement them, enabling iterative development with large language model agents.

SSRLive: Live Streaming Recommendation with Dynamic Semantic ID
Graph Learning

SSRLive proposes live streaming recommendation with Dynamic Semantic IDs to improve real-time adaptation and resource utilization. Leveraging generative recommendation, it adapts to changing streams and user interactions more efficiently.

HKVM-RAG: Key-Value-Separated Hypergraph Evidence Organization for Multi-Hop RAG
Graph Theory

HKVM-RAG introduces a key-value-separated evidence-organization layer for multi-hop RAG. It assembles answer-path hyperedges from cached passage-level evidence tuples, organizing evidence into hypergraph structures to enhance retrieval fidelity.

HiPS: Hierarchical PDF Segmentation of Doctrinal Legal Books
Knowledge Graph LLM × Graph

HiPS tackles hierarchical PDF segmentation of doctrinal legal books, going beyond page-level layout understanding to recover deep section hierarchies. It introduces a method and releases gold-standard data to support training and evaluation of deep document hierarchies.