← Home

Daily arXiv Papers

Graph Neural Networks · Graph Learning · LLM × Graph

Showing 21 papers for 2026-02-23

COMBA: Cross Batch Aggregation for Learning Large Graphs with Context Gating State Space Models
Graph Learning

COMBA proposes Cross Batch Aggregation for learning large graphs with state space models. It introduces graph-context gating to efficiently process long-range dependencies without converting entire graphs into sequences.

Causal Neighbourhood Learning for Invariant Graph Representations
Graph Learning

The paper advocates causal neighbourhood learning to obtain invariant graph representations under distribution shifts. It aims to suppress spurious correlations and improve generalization across graphs.

Optimizing Graph Causal Classification Models: Estimating Causal Effects and Addressing Confounders
Graph Learning

This work focuses on optimizing graph causal classification models by estimating causal effects and addressing confounders in graph-structured data. It discusses methods to identify causal relationships rather than mere associations.

Generating adversarial inputs for a graph neural network model of AC power flow
GNN

The paper formulates optimization problems to generate adversarial inputs for a graph neural network predicting AC power flow. Demonstrated on a 14-bus CANOS-PF GNN, adversarial points yield large errors. The study also discusses the required perturbation magnitude and robustness.

Advection-Diffusion on Graphs: A Bakry-Emery Laplacian for Spectral Graph Neural Networks
GNN

Introduces Bakry-Emery graph Laplacian combining diffusion and advection via a learnable node potential, enabling task-specific propagation without changing graph topology. It yields a well-behaved spectral decomposition and can replace standard Laplacians in spectral GNNs.

Stable Long-Horizon Spatiotemporal Prediction on Meshes Using Latent Multiscale Recurrent Graph Neural Networks
GNN

Proposes a deep learning framework for stable long-horizon spatiotemporal prediction on meshes, using latent multiscale recurrent graph neural networks to predict full temperature histories. The approach aims to improve accuracy and stability while controlling computational costs.

Parameter-Efficient Domain Adaptation of Physics-Informed Self-Attention based GNNs for AC Power Flow Prediction
GNN

Studies parameter-efficient domain adaptation for physics-informed self-attention GNNs in AC power flow prediction under domain shift from MV to HV grids. It seeks to retain source-domain knowledge while adapting to target domains with limited retraining.

Unifying approach to uniform expressivity of graph neural networks
GNN

Unifies expressivity of GNNs by introducing Template GNNs (T-GNNs), a framework that extends beyond WL by incorporating substructure information via templates. The approach aims to unify various expressive architectures.

CityGuard: Graph-Aware Private Descriptors for Bias-Resilient Identity Search Across Urban Cameras
Graph Learning

CityGuard presents a graph-aware private descriptor approach for privacy-preserving identity search across urban cameras. It uses a dispersion-adaptive metric learner and spatially conditioned attention within a topology-aware transformer to reduce bias and protect privacy.

Benchmarking Graph Neural Networks in Solving Hard Constraint Satisfaction Problems
GNN

Benchmarking Graph Neural Networks in Solving Hard Constraint Satisfaction Problems. The paper introduces new hard benchmarks for CSPs based on random problems and compares GNNs to classical heuristics. It finds classical algorithms still outperform GNNs and discusses challenges for neural approaches.

Co-Evolution-Based Metal-Binding Residue Prediction with Graph Neural Networks
GNN

Co-Evolution-Based Metal-Binding Residue Prediction with Graph Neural Networks. The paper uses GNNs to predict metal-binding residues and metal types by incorporating co-evolutionary constraints. It argues this captures evolutionary dependencies that traditional sequence- and structure-based methods miss.

Beyond Simple Graphs: Neural Multi-Objective Routing on Multigraphs
Graph Learning

Beyond Simple Graphs: Neural Multi-Objective Routing on Multigraphs. It proposes GNN-based routing methods for multigraphs with multiple edges per node pair. The methods autoregressively select edges to optimize multiple objectives, showing improved routing performance.

Physics-informed GNN for medium-high voltage AC power flow with edge-aware attention and line search correction operator
GNN

Physics-informed GNN for medium-high voltage AC power flow with edge-aware attention and line search correction operator. It presents PIGNN-Attn-LS, combining edge-aware attention that encodes line physics with per-edge biases, and a line-search correction to improve inference accuracy. The approach enhances robustness across operating scenarios.

Physics-informed graph neural networks for flow field estimation in carotid arteries
GNN

Physics-informed graph neural networks for flow field estimation in carotid arteries. Proposes physics-informed GNNs to estimate hemodynamic flow fields in carotid arteries using symmetry and physical priors, enabling accurate predictions from limited data. The method aims to support non-invasive assessment of cardiovascular risk.

ViGText: Deepfake Image Detection with Vision-Language Model Explanations and Graph Neural Networks
GNN

ViGText: Deepfake Image Detection with Vision-Language Model Explanations and Graph Neural Networks. Introduces ViGText, which integrates Vision-Language Model explanations with graph neural networks to detect deepfakes. The cross-modal explanations improve detection performance and robustness to attacks.

Ensemble-based graph representation of fMRI data for cognitive brain state classification
Graph Learning

Ensemble-based graph representation of fMRI data for cognitive brain state classification. Proposes an ensemble-based graph representation where edge weights encode state evidence via posterior-probability differences from edge-wise classifiers. This ensemble approach is evaluated on cognitive state datasets.

LLM-WikiRace Benchmark: How Far Can LLMs Plan over Real-World Knowledge Graphs?
Knowledge Graph

LLM-WikiRace Benchmark: How Far Can LLMs Plan over Real-World Knowledge Graphs? Introduces LLM-Wikirace, a benchmark to test planning and reasoning over real-world knowledge graphs by navigating Wikipedia hyperlinks. The study evaluates several models and reveals their planning strengths and limits.

El Agente Gr\'afico: Structured Execution Graphs for Scientific Agents
Graph Learning

El Agente Gráfico: Structured Execution Graphs for Scientific Agents. Presents El Agente Gráfico, a framework that embeds LLM-driven decision making inside a type-safe execution graph to improve control, provenance, and auditability of scientific workflows.

Enhancing Scientific Literature Chatbots with Retrieval-Augmented Generation: A Performance Evaluation of Vector and Graph-Based Systems
Graph Learning

Enhancing Scientific Literature Chatbots with Retrieval-Augmented Generation: A Performance Evaluation of Vector and Graph-Based Systems. Evaluates RAG-based literature chatbots using both vector and graph-based retrieval, across scenarios of single-document and large-scale corpus retrieval. The results discuss trade-offs in speed, coverage, and accuracy.

CUICurate: A GraphRAG-based Framework for Automated Clinical Concept Curation for NLP applications
LLM × Graph Graph Learning

CUICurate: A GraphRAG-based Framework for Automated Clinical Concept Curation for NLP applications. Introduces CUICurate, a GraphRAG framework to automatically curate clinical concept sets that group synonyms, subtypes, and supertypes for NLP tasks. It targets UMLS CUIs and improves accuracy and efficiency of concept-set construction.

Beyond Homophily: Community Search on Heterophilic Graphs
Graph Theory

This paper studies community search on heterophilic graphs, where edges predominantly connect dissimilar nodes, in contrast to traditional homophily assumptions. It argues that structural signals in such graphs manifest as sharp high-frequency contrasts, causing classical methods like k-core and k-truss—as well as many ML-based models—to struggle to identify query-relevant node sets. To address this, the authors propose a framework tailored to heterophilic networks that accurately extracts refined communities centered around a given query.