Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 37 papers for 2026-08-18
BDIP-Net introduces a dual-interaction graph learning framework for bilayer materials that explicitly models intra-layer covalent bonds and inter-layer van der Waals interactions to predict properties without relying on expensive DFT optimizations. By constructing two interaction graphs and jointly learning from them, it improves accuracy and transferability across stacking configurations.
This work disentangles whether failures of graph neural networks on heterophilic nodes are due to heterophily itself or the rarity of such nodes. Through systematic evaluation of six GNNs on five datasets with varying homophily, the study shows that homophilic nodes are generally easier to classify, while rarity exacerbates errors for many models.
M-LINKX proposes Multiview Graph Learning for brain cognitive disease detection from EEG data. It handles long EEG recordings by segmenting them into fixed-length windows and learns from multiple segment views to improve discrimination among Alzheimer's disease, mild cognitive impairment, and frontotemporal dementia. The multiview approach enhances robustness to inter-subject variability and noise.
GATTA introduces Graph Active Learning with Test-Time Augmentation, a framework that improves active learning on graphs by aggregating predictions across multiple augmented views to yield more reliable uncertainty estimates. It includes a consistency-based filtering mechanism to handle label-preserving graph augmentations.
FinFraudBench is a heterogeneous-graph benchmark for financial fraud detection. It addresses misalignment between existing public benchmarks and real-world financial systems by constructing realistic relational datasets among customers, cards, merchants, categories, and locations to enable relational risk reasoning.
Amortised Post-Hoc Explanation with Exact Preservation for Dynamic Graph Anomaly Detectors introduces X-StrGNN, a post-hoc explanation layer that wraps a frozen StrGNN anomaly detector and emits attribution vectors for every flagged edge, providing amortised explanations with exact preservation guarantees.
NICE: Scale-Stable Perturbations for Graph Neural Network Explanations via Noise Corruption proposes a perturbation-based explanation method that uses noise corruption to stabilize explanations and reduce distribution shift, addressing limitations of Element-wise Masking.
Unifying Graph Neural Networks Through a Common Layer Equation presents a common layer equation that represents GNNs with seven components, factorizing where information moves (propagation) from what moves (messages). This unified view reveals shared computations across architectures.
Multi-Feature Riemannian Hypergraph for Online Test-Time Adaptation of Motor Imagery Brain-Computer Interface (MRieHy) leverages hypergraphs and Riemannian geometry to improve online test-time adaptation and cross-day transfer in MI-BCI decoding.
Demystifying Oversmoothing in Sheaf Neural Networks: An Index-Theoretic Criterion argues that the harmonic space dimension alone is insufficient to characterize oversmoothing in SNNs. It introduces an index-theoretic criterion that better captures anti-oversmoothing capacity and provides theoretical insights.
Longitudinal and Graph-Augmented Prediction of Adolescent Substance Use Onset in the ABCD Study compares cross-sectional, longitudinal, and graph-based approaches to predicting alcohol sipping, alcohol use, marijuana use, and combined alcohol/marijuana use in about 11,860 participants, highlighting the added value of relational information.
BRAID: Learning Equilibrium Maps in Interdependent Security Games via Weight-Tied Iterative Graph Neural Networks introduces BRAID, a weight-tied iterative GNN that directly maps game parameters to Nash equilibrium effort profiles, replacing slow best-response dynamics for fast equilibrium prediction.
BrainLinear: A Linear Model for Brain Network Analysis in Sparse Tangent Subspaces proposes BrainLinear, a linear model operating in sparse tangent subspaces of the functional connectome to reduce computation and improve interpretability while preserving predictive power.
Domain-Agnostic Neural Topic Modeling with Contextual Token-Level Semantic Graph Representation develops domain-agnostic neural topic modeling that uses contextual token-level semantic graphs to improve topic interpretability across domains, addressing domain-specific term misrepresentation in pre-trained embeddings.
POI Recommendation with LLM-Augmented Multi-Graph Learning and Contrastive Alignment introduces LLM-MGCL, a multi-graph recommender that leverages semantic information from large language models and two auxiliary item-item graphs (semantic and spatial), with contrastive alignment to improve cold-start POI recommendations.
Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing offers an explainable framework that routes anomalies to specialized experts corresponding to different causal mechanisms (price shocks, liquidity events, contagion, etc.) and provides mechanism-level explanations.
Not All Neighbors Matter: Understanding the Impact of Graph Sparsification on GNN Pipelines investigates how graph sparsification affects GNN performance at scale, showing that not all neighbors contribute equally and offering practical guidelines for when and how to sparsify.
Learning Optimal Dynamic Matching via Graph Neural Networks develops a value-based reinforcement-learning framework for dynamic matching on finite evolving weighted graphs, proving an event-time reduction: the planner can act immediately after exogenous events without loss of optimality.
EquiPocket: an E(3)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction uses an E(3)-equivariant GNN to predict protein ligand binding sites, addressing rotation sensitivity and irregular protein structures with robust surface-aware representations.
Transferable FB-GNN-MBE Framework for Potential Energy Surfaces: Data-Adaptive Transfer Learning in Deep Learned Many-Body Expansion Theory presents FB-GNN-MBE, integrating a fragment-based GNN into many-body expansion theory to reproduce first-principles potential energy surfaces across larger systems via data-adaptive transfer learning.
TAHB introduces the Text-Attributed Hypergraph Benchmark, the first public benchmark that combines hypergraph structures with raw textual attributes. It enables evaluation of text-aware hypergraph learning by linking language-model style semantics with higher-order relations, and it defines datasets, tasks, and metrics to support research progress.
GraphLoom proposes a reliability-calibrated multimodal KG-RAG framework for faithful evidence routing. It builds an instance-level multimodal knowledge graph from grounded scene descriptions and relational extractions, and uses reliability calibration to route compact, trustworthy evidence for generation. The result is more faithful and efficient multimodal retrieval-augmented generation.
Grounding Healthcare LLMs in a Causal Knowledge Graph outlines a reproducible, graph-centered evaluation framework for intervention-oriented healthcare reasoning, and stress-tests it in a cardiovascular pilot. The framework treats causal knowledge assertions as first-class nodes with provenance and supports evaluation across interventions, mechanisms, harms, evidence, and uncertainty. A cardiovascular pilot demonstrates how such evaluation can diagnose LLM reasoning gaps and guide improvements.
Schema-Agnostic Graph Reasoning Agent (GRA) introduces a graph-based reasoning agent that treats hybrid knowledge graphs (textual concepts and relational tables) as a common interface. It provides seven generic tools to list neighbours, read node content, and search descriptions, enabling domain-agnostic exploration of unfamiliar codebases and data schemas. The agent demonstrates effective reasoning by discovering domain-specific structures through graph operations.
Hypergraph-based Multimodal Retrieval-Augmented Generation with Incremental Refinement presents a framework that uses hypergraphs to encode high-order relations across modalities (e.g., charts, text, numerical data). It replaces full-page reconstructions with incremental refinement to reduce computation while preserving cross-modal alignment. Experiments show improved retrieval quality and generation fidelity over traditional graph-based RAG.
Quipu introduces a governable, embeddable bitemporal knowledge graph store that inverts common defaults in write governance. It gates every fact through predicates evaluating the pending post-state, and records data, trust labels, verdicts, and governance metadata to support agent workloads. This design supports robust provenance and governance in dynamic KGs.
Valhalla proposes a layered knowledge-state and service-governance framework to sustain long-term scientific knowledge work with LLM agents. It shifts beyond node-centric graphs toward layered memory, external knowledge bases, and governance services to enable knowledge sharing, reorganization, and task continuity across users and projects. The framework supports persistent provenance, reuse, and governance across the scientific lifecycle.
Characterising cardiac tissue properties with graph neural networks develops a GNN-based framework trained on synthetic electrogram signals to identify regions of interest for premature ventricular complex ablation. The method achieves high average precision around 0.95–0.97 in detecting target areas on 2D representations, demonstrating potential to aid targeted ablations.
Noesis presents a decoupled Graph-RAG architecture with bidirectional retrieval, adaptive parallelism, and cross-knowledge-base semantic discovery. It addresses long-document ingestion, retrieval scalability, and cross-domain deployments by separating components and enabling flexible routing and grounding. The result is scalable, multi-domain graph-grounded LLM applications.
Graph Neural Network-assisted A2C optimizes latency-sensitive edge vision systems such as UAVs by using a GCN-enhanced actor-critic DRL to make transmission and computation decisions under strict latency constraints. The approach reduces end-to-end latency while maintaining or improving perception performance. Experiments demonstrate gains over baseline methods.
LlamaRec-LKG-RAG introduces a single-pass, learnable Graph-RAG framework that integrates personalized knowledge-graph context into LLM-based ranking for recommendations. The architecture extends LlamaRec with an LKG component to enable end-to-end training and improved ranking quality. The results show more accurate and context-aware recommendations.
A large-scale Chinese KG-text alignment dataset (CDTP) is introduced to benchmark knowledge-grounded LLMs. CDTP contains over 7 million aligned instances across four domains, enabling robust evaluation of Chinese KG to text reasoning and factual grounding. The dataset supports diverse alignment types and downstream tasks.
This study measures cross-boundary collaboration in an OSS ecosystem by reconstructing a large bipartite contributor-repository graph for 464 cybersecurity projects and 11,372 contributors. It uses Louvain community detection to identify 163 non-singleton communities and finds superlinear scaling of contributors with repositories; recognition and repeat collaboration patterns sustain cross-community work. The findings offer governance implications for open source ecosystems.
LineageRAG constructs evidence lineages for each query-derived evidence demand and expands them with demand-conditioned retrieval, then appends verbatim source spans to satisfy the lineage. It unifies evidence discovery and source grounding, improving traceability and transparency in GraphRAG. The approach supports multi-hop reasoning with explicit source traces.
A graph-based framework uncovers mathematical software communities from a software co-usage network built from publications and swMATH; community detection reveals heterogeneous software ecosystems, and the approach predicts publication-to-community associations. The method demonstrates cross-disciplinary software use and helps map how software tools relate to research outputs.
GALLM (Graph-Aware LLMs for Sequential Recommendation) presents a graph-aware LLM framework that integrates collaborative signals directly into the language model, avoiding external recommender modules. It models global collaborative patterns beyond intra-sequence dependencies and demonstrates improved sequential recommendation performance. The approach achieves better personalization and robustness.
Evidence-Carrying Validation for Knowledge Graphs introduces an evidence-carrying validation interface for schema-based checks. Each selected node-shape check returns not only pass/fail but supporting evidence and partial-match explanations, enabling explainable validation and easier debugging for KG consumers. The design enhances trust and traceability in graph validation.