Showing 22 papers for 2026-05-18
Mask-Morph Graph U-Net is proposed as a generalizable mesh-based surrogate for crashworthiness field prediction under large geometric variation. It combines a mask-morphing mechanism with a graph U-Net to capture nonlinear relationships across diverse mesh topologies. The model aims to enable faster iterative design optimisation while maintaining accuracy.
Njord is a probabilistic data-driven ocean forecasting model that blends a deep latent variable framework with a graph neural network. This enables sampling of each forecast step in a single forward pass, producing stochastic predictions. It demonstrates global and regional applications across different resolutions, enabling scalable ensemble ocean forecasts.
OgBench introduces a framework to evaluate Graph Neural Networks on omics data, addressing the n << p regime typical of biology. It emphasizes low-sample, high-node graphs such as genes, transcripts, or proteins across few patients. The benchmark helps quantify GNN performance in this challenging domain.
GelGT introduces Gaussian Relational Graph Transformer, addressing long-range dependencies in relational graphs and jointly modeling structural, semantic, and temporal information. The model uses Gaussian relational mechanisms to better capture uncertainty and interactions across distant nodes. This leads to improved relational reasoning on diverse tasks.
Context-aware entity-relation extraction is proposed for Cybersecurity Knowledge Graphs (CKGs). It tackles the challenge of extracting triples from unstructured CTI reports, where complex report structure and domain-specific language cause ambiguity. The approach improves extraction robustness by leveraging contextual cues.
We diagnose attention dispersion as a common failure mode for CTDG transformers under temporal distribution shift. Through controlled ablations, we show predictions depend on a small set of historically predictive nodes. We also propose a transferable fix that improves robustness across datasets.
We present an inductive approach for applying gradient-boosted decision trees on graphs to insurance fraud detection. The method addresses class imbalance and heterogeneity in dynamic, multi-relational graphs common to insurance data. Empirical results demonstrate improved detection performance.
Graph-Regularized Sparse Autoencoders (GSAE) are proposed to learn safety-steering directions for LLMs. By smoothing decoder vectors over a neuron co-activation graph, GSAE captures distributed activation patterns that relate to safe or harmful outputs. This improves steering decisions during inference.
T2T-LA studies topology-to-topology learning for graph problems without feature access or task knowledge. It leverages LLMs to translate topology information into actionable guidance, enabling effective graph learning in settings where features and labels are unavailable.
Flowette introduces a continuous flow-matching framework for graph generation, using a graph neural network transformer to learn a velocity field over graph representations. It promotes topology-aware alignment via optimal transport and incorporates graphette priors for domain-specific structure.
The paper combines a node-transformer architecture with BERT-based sentiment analysis to predict stock market movements. The integration helps capture cross-sectional dependencies and market sentiment signals, improving forecasting performance.
Do Larger Models Win in Drug Discovery? The study benchmarks model scaling across 26 endpoints for molecular property, toxicity, and bioactivity prediction. It analyzes whether larger pretrained models consistently outperform smaller, task-specific GNNs across diverse datasets.
AIMing for Standardised Explainability Evaluation in GNNs proposes a framework for evaluating explainability, focusing on graph kernel networks, and provides guarantees such as complete coverage and contamination resistance.
Building Specialized Software-Assistant ChatBot with Graph-Based Retrieval-Augmented Generation proposes software-assistant chatbots guided by graph-based retrieval augmented generation to reduce hallucinations and improve reliability.
Towards Foundation Models for Relational Databases with Language Models and Graph Neural Networks discusses creating foundational models that operate on relational databases using relational entity graphs (REGs) and combines LMs with GNNs for generalized relational learning.
From Guidelines to Guarantees presents a graph-based evaluation harness that converts structured clinical guidelines into a knowledge graph to dynamically instantiate evaluation queries, ensuring comprehensive coverage and resistance to data contamination.
Falkor-IRAC argues for graph-constrained generation to support verified legal reasoning in Indian judicial AI, addressing limitations of vector-based RAG in legal contexts. It demonstrates constrained symbolic reasoning for precedents and statutes.
Region-Grounded Report Generation for 3D Medical Imaging introduces VietPET-RoI, a large-scale 3D PET/CT dataset with fine-grained RoI annotations for low-resource languages, and a graph-enhanced framework for region-grounded report generation.
ShadowMerge reveals a poisoning attack on graph-based agent memory by injecting crafted relation-channel conflicts, causing retrieved memory to bias agent behavior. The paper analyzes vulnerability and proposes defensive considerations.
Representing Higher-Order Networks surveys graph-based frameworks that extend beyond pairwise interactions, including multiway, temporal, multilayer, recursive, and tensor-based models, providing a comprehensive taxonomy and discussion of applications.
This work presents a generalizable method that leverages large language models together with open web data to reconstruct the temporal, multi-relational firm network of the semiconductor sector at scale. It tackles major drawbacks of proprietary databases—high cost, incompleteness, and slow updates—by enabling rapid, scalable extraction of inter-firm relationships. The approach captures the evolving dependencies and structural dynamics of the industry, with implications for researchers, firms, and policymakers.
The authors automatically construct the first large-scale legal citation graph from 100.7 million Ukrainian court decisions by extracting about 502 million citation links across six types using regex on commodity hardware, in roughly five hours. They show that judicial citation structure encodes legal-domain boundaries and can predict future legislative importance with near-perfect accuracy, enabling topological analysis and ontology-driven clustering. The workflow scales to about a terabyte of data, demonstrating feasible, unsupervised extraction for massive legal corpora.