Showing 27 papers for 2026-05-27
We propose CE-FedGNN, a framework for learning on coupled graphs in a federated setting where raw data cannot be shared. The method achieves communication efficiency and privacy guarantees while preserving cross-client linkage information without exchanging dense embeddings, enabling accurate GNN training across participants.
We address dynamic link prediction in temporal signed networks by introducing a modular temporal enhancement for signed GNNs. Our approach captures the evolution of both positive and negative relations while respecting balance-theoretic constraints, improving predictive performance on evolving graphs.
We introduce ProMoS, the first unsupervised generalist GAD framework that learns a prototype-based distillation to produce transferable anomaly detectors across graphs. It reduces dependence on scarce annotations and avoids few-shot inference at test time, improving robustness to unseen anomaly patterns.
TED models tax evasion detection as a heterogeneous graph task guided by related party transactions. It exploits rich interactive information in tax scenarios beyond company-level statistics. The framework achieves improved detection accuracy on heterogeneous graphs.
This study asks what pharmacology can be inferred from molecular structure in toxicity prediction. Through a systematic case study centered on aspirin, we analyze explainability gaps in GNN-based toxicity models and taxonomy of what the structure cannot encode.
We propose using a knowledge graph as a missing data layer for LLM-based industrial asset operations. A KG-backed data layer improves reasoning beyond flat documents, demonstrated on scenarios with a 781-node graph built from CouchDB/YAML/CSV data. We compare data-layer choices and show gains.
We present BSMS-GNN for efficient mesh-based physical simulation with multi-scale GNNs to address scaling and oversmoothing. It reduces reliance on manual coarse meshes by learning scalable multi-scale structures, enabling accurate simulations on large meshes.
We introduce Graph State-Space Models (GSSMs) that embed state-space computation directly into message-passing neural networks. This preserves permutation equivariance and efficiency while improving modeling of long-range temporal dependencies on graphs.
LGKDE learns kernel density estimation for graphs by mapping each graph with a GNN into a distribution and learning the KDE. It combines graph kernels with KDE under learnable representations to improve graph-level anomaly detection.
Morphling delivers fast, fused, and scalable GNN training by designing specialized kernels that fuse irregular graph traversals with regular dense matmul. This reduces memory movement and intermediate allocations, enabling efficient training at scale.
We study MPLang, a declarative language for GNN-style computation, and characterize the expressive power of A-MPLang (without activations) via walk-summed features. For bounded activations, we show that all eventually constant activations yield the same numerical and Boolean expressive power, unifying prior results.
GraphDancer is a two-stage curriculum post-training framework that teaches LLMs to reason over graphs by interleaving natural language reasoning with graph function execution. It enables LLMs to explore graphs and perform precise graph-aware reasoning.
This work introduces SVBCX for silicosis/pneumonia CXR classification and proposes a graph Transformer post-hoc modeling with ensemble techniques to boost performance.
We study distributed control of networked systems by parameterizing GNN policies within a Youla-like framework, ensuring stability by design. The magnitude component uses a stable operator with a GNN on disturbances, while the direction component uses a GNN.
NeuroMambaLLM combines dynamic graph learning of fMRI functional connectivity using Mamba with language model reasoning to support autism neuroscience analysis. It models transient FC dynamics and leverages LLMs to interpret results.
PhyGHT introduces Physics-Guided HyperGraph Transformer to purify signals at HL-LHC by modeling pileup as a hypergraph with physics priors, improving signal extraction.
GraphIP-Bench provides a unified benchmark to evaluate GNN model extraction attacks and defenses under consistent threat models, enabling fair comparisons of theft resistance and defense effectiveness.
Helicase presents uncertainty-guided, autonomous multi-agent LLMs to construct supply chain knowledge graphs from fragmented web sources via multi-hop reasoning, enabling reliable decision support.
We propose enhanced negative sampling methods to train KG foundation models, improving zero-shot KG completion on unseen KGs with different relation vocabularies. Empirical results show better generalization.
Chat-ISV is a KG-enhanced multi-agent Q&A system for steel industry VOC governance. It constructs a Neo4j knowledge graph from literature (27180 nodes, 81779 relations) to provide traceable reasoning and reliable decision support.
This study examines how different facts in knowledge graphs influence hypothesis generation for battery materials across multiple LLMs (Mistral-7B, Llama-3.1-70B, Gemini 2.5 Flash). The authors perturb local KGs by varying density, ontology richness, topology, and control structure, and assess outputs with both provided-graph and fixed-reference metrics. They find that KG utility is selective and highly model-dependent: adding graph context changes outputs, but there is no universal mapping from a graph property to improved hypothesis quality.
RAGEAR is a neurosymbolic recommender for academic courses that combines dense retrieval over full lecture transcripts with a symbolic Knowledge Graph capturing courses, lessons, transcript chunks, credits, study plans, and curricular constraints. The Knowledge Graph enables symbolic filtering and contextualization based on constraints such as credits, disciplines, study plans, and prerequisites, delivering more controllable and explainable recommendations than metadata-only baselines.
OMD-GraphRAG extends GraphRAG with three core innovations: Ontology-Guided Knowledge Extraction using predefined ontologies or schemas to improve extraction precision; Multi-Dimensional Clustering to organize retrieved knowledge across modalities and queries; and Dual-Channel Fusion that combines graph-structured signals with text-based signals for better retrieval and reasoning. The framework targets addressing challenges in knowledge extraction precision, community report integrity, and retrieval performance in complex reasoning tasks.
Self-signals Driven Multi-LLM Debate for Efficient and Accurate Reasoning introduces a self-signal-guided multi-LLM debate (MAD) approach that uses internal generation signals (e.g., token logits and attention) to steer discussions, reducing redundancy from relying solely on external debate graphs or judges. This self-signal integration improves efficiency and accuracy in multi-agent reasoning.
SEAL: Self-Evolving Agentic Learning for Conversational Question Answering over Knowledge Graphs presents a self-evolving agentic learning framework to improve knowledge-graph-based conversational QA. It addresses coreference resolution, contextual dependency, and complex logical reasoning while reducing inaccuracies and computational costs associated with generating and parsing complex logical forms.
MuNet proposes a mutualistic network for joint 3D human mesh recovery and 3D clothed human reconstruction from single images, enabling the two related tasks to benefit from shared representations. The approach uses 2-manifold graphs to model geometric relations and perform joint optimization, yielding improvements over separate-task baselines.
Rethinking Agentic RAG argues for moving beyond embedding-based retrieval toward LLM-driven logical retrieval. It emphasizes that agentic RAG should enable LLMs to construct structured, logical queries and reason over retrieved results, rather than relying solely on dense embeddings or back-end retrieval complexity. The paper outlines a framework for delegating query construction to the LLM to improve retrieval efficiency and reasoning quality.