Showing 17 papers for 2026-03-25
We propose two data-driven chemical mechanism reduction formulations based on graph neural networks with message-passing transformer layers to learn nonlinear dependencies among species and reactions, addressing the computational burden of direct numerical simulations of detailed mechanisms. The first formulation, GNN-SM, uses a pre-trained surrogate model to guide reduction across broad conditions, while the second provides an additional reduction strategy leveraging learned dependencies.
This work applies conformal risk control (CRC) to wildfire spread prediction, providing distribution-free safety guarantees on missed fire spread inferences to support evacuation planning. It compares tabular, spatial, and graph-based models under CRC, delivering formal finite-sample risk bounds for decision-making in safety-critical contexts.
We propose Symbolic Graph Network (SGN) for PDE discovery under noisy and sparse data. By avoiding local differential approximations and leveraging symbolic graph representations, SGN learns governing equations robustly in challenging data regimes and captures underlying physical laws.
MLANet is a universal and efficient graph neural network framework for machine learning interatomic potentials, featuring a dual-path dynamic attention mechanism. It aims to deliver near-quantum accuracy at linear computational cost while maintaining stability across diverse atomic environments.
This study investigates whether graph foundation models (GFMs) can generalize across architectures. It finds that true zero-shot generalization is constrained by reliance on a fixed architectural backbone, highlighting the need for architectural diversity to enable broader cross-graph generalization.
Q-AGNN introduces a Quantum-Enhanced Attentive Graph Neural Network for intrusion detection, modeling network flows as nodes and similarity relations as edges. The quantum-inspired attention mechanism enhances the ability to capture relational dependencies for more effective threat detection.
HGNet presents a scalable foundation model for automated knowledge graph generation from scientific literature, addressing long multi-word entity recognition, cross-domain generalization, and hierarchical knowledge representation. It demonstrates improvements in coverage and consistency over baselines.
Graph Variate Neural Networks (GVN) address dynamically evolving spatio-temporal signals on graphs by constructing a network tensor of instantaneous connectivity profiles against a stable support. This framework enables robust modeling of changing graph structures over time.
MemReward introduces a graph-based experience memory to predict rewards for LLM RLHF when labeled data are scarce. The memory stores experiences as graphs and reuses them to improve reward estimation and sample efficiency.
MoEGCL proposes Mixture of Ego-Graphs Contrastive Representation Learning for multi-view clustering, constructing ego-graphs per view and learning cross-view representations to achieve more effective graph fusion and clustering.
Dynamic Fusion-Aware Graph Convolutional Network (DF-GCN) for Multimodal Emotion Recognition in Conversations models MERC with a fusion mechanism that dynamically adapts across emotion types, leveraging speaker relationships and modality interactions to improve accuracy.
GraLC-RAG introduces Graph-Aware Late Chunking for Retrieval-Augmented Generation in biomedical literature, unifying late chunking with graph-aware structural intelligence to surface evidence across document sections and improve evidence retrieval.
CLiGNet is a Clinical Label-Interaction Graph Network for medical specialty classification from clinical transcriptions. It identifies severe data leakage from SMOTE pre-splitting, establishes a leakage-free benchmark across 40 specialties, and uses a label-interaction graph to improve classification.
GraphRAG for Engineering Diagrams: ChatP&ID enables LLM interaction with P&IDs via grounded graph retrieval-augmented generation, reducing processing costs and hallucinations when handling engineering diagrams.
Curriculum-Driven 3D CT Report Generation uses a four-phase curriculum and language-free visual grafting with zone-constrained compression to generate radiology reports from 3D CT volumes, leveraging a frozen visual encoder and a Llama 3.2 decoder for grounded outputs.
Towards Intelligent Geospatial Data Discovery proposes a knowledge graph-driven multi-agent framework powered by large language models to handle distributed, heterogeneous geospatial data and improve discovery through graph-based reasoning.
MS-DGCNN++ introduces multi-scale dynamic graph convolution with scale-dependent normalization to robust LiDAR tree species classification, addressing density variations between trunk and canopy. Normalization reduces scale-related bias, improving classification accuracy.