Showing 28 papers for 2026-05-15
GFMate proposes test-time prompt tuning for Graph Foundation Models to adapt to new graphs with minimal labeled data. By decoupling source-domain information from prompts and enabling prompt-only adaptation at inference, it reduces entanglement with pretraining and improves few-shot performance.
This work investigates leveraging hidden contextual information for forecasting dynamic movements, tracing a journey from recurrent networks to graph neural networks and general-purpose transformers to better capture inter-player dynamics. It contrasts traditional time-series methods (ARIMA, Kalman, Particle filters) with modern graph- and transformer-based models to handle non-linear dynamics and abrupt changes in trajectories.
AIMing for Standardised Explainability Evaluation in GNNs introduces a standardized framework for evaluating explainability in Graph Neural Networks, enabling apples-to-apples comparison across models and methods. The paper includes a case study on Graph Kernel Networks to demonstrate reproducible, comparable explainability assessments.
We introduce ASTDP-GAD, a neuromorphic approach to graph anomaly detection that fuses spiking graph neural networks with adaptive STDP learning for energy-efficient, time-aware anomaly detection in dynamic networks. The framework unifies spiking computation, STDP-based learning, and graph-based anomaly detection, offering robustness and low power consumption.
LEAP proposes Local Euler Characteristic Transform (ECT)-based learnable positional encodings for graphs to enrich node representations with topological signals. By providing differentiable, efficiently computable encodings, LEAP enhances message passing networks beyond standard local neighborhoods.
Entity alignment with graph foundation models faces a reasoning horizon gap: long-range dependencies across sparse, heterogeneous knowledge graphs are hard to capture. The paper proposes a foundation-model-driven EA framework to extend reasoning beyond local neighborhoods and improve alignment across unseen graphs.
Hyperbolic Graph Neural Networks are challenged by geometry-task alignment: the paper argues that the success of HGNNs depends on whether the metric geometry aligns with the task's inherent structure. Through theory and experiments, it shows that proper alignment yields better representations, while misalignment can hinder performance.
Contraction and Hourglass Persistence for Learning on Graphs, Simplices, and Cells introduces Contraction Homology as a topological learning tool, addressing limitations of persistent homology's inclusion-based workflows. It analyzes contraction sequences and their persistence, offering a robust alternative for learning on graphs and simplicial complexes.
PRAETORIAN targets the intrinsic requirements of effective GNN backdoors rather than surface cues. By analyzing internal correlations and external triggers, the defense aims to detect and mitigate backdoors that rely on substantial influence to flip a victim node's prediction, offering robustness against adaptive attackers.
DRIFT proposes a benchmark for task-free continual graph learning under continuous distribution shifts, reflecting realistic non-stationary environments. It provides protocols, baselines, and evaluation metrics to assess how models adapt while mitigating catastrophic forgetting.
TIDE presents Tri-Component Information Decomposition for graph OOD detection, separating information into feature and structure components and their interactions. The framework helps identify which components matter for robustness, improving detection under distribution shifts.
MLGIB introduces a Multi-Label Graph Information Bottleneck to address over-squashing in deep message passing on multi-label graphs. By applying an information bottleneck, it preserves discriminative label signals while discarding irrelevant or noisy information.
NeuroMambaLLM combines dynamic graph learning of fMRI functional connectivity using Mamba with language-model reasoning to analyze autistic brains. It captures transient connectivity patterns and leverages LLM reasoning to interpret brain-network dynamics.
Strikingness-Aware Evaluation for Temporal Knowledge Graph Reasoning introduces a strikingness measure to TKGR evaluation, using a rule-based framework (RSMF) to quantify how rare or高难度的事件的预测需要更深的推理。该框架强调对具有高推理难度事件的评估。
IdeaForge presents a knowledge graph-grounded multi-agent framework for cross-methodology innovation analysis and patent claim generation, integrating multiple ideation methodologies (e.g., TRIZ, Design Thinking) with reasoning agents and a knowledge graph to preserve intermediate reasoning and support novel idea synthesis.
Representing Higher-Order Networks surveys graph-based frameworks that capture multiway, hierarchical, temporal, multilayer, recursive, and tensorial interactions, offering a comprehensive account of higher-order representations beyond pairwise edges.
Beyond Individual Mimicry introduces GraphMind, a framework that augments LLM-driven social agents with graph structure to enable more human-like global interactions. The approach enhances realism and resilience to graph-based bot detection by coordinating agents over a connected social graph.
Towards Robust Federated Multimodal Graph Learning studies federated learning for multimodal graphs with modality heterogeneity and incomplete data. It proposes robust representation alignment and aggregation techniques to handle missing modalities and cross-party data constraints.
GraphIP-Bench presents a unified benchmark for evaluating GNN model-extraction attacks and defenses, enabling consistent comparisons across datasets, threat models, and metrics. It highlights the practicality of assessing theft risk and defense effectiveness in real-world cloud deployments.
Rethinking Efficient Graph Coarsening promotes a non-selfishness principle for graph coarsening, shifting from node-by-node selfish partner matching to cooperative, global partner selection. This reduces computational and memory overhead while preserving essential structural and semantic properties.
Dynamic graphs from different domains are hard to model with a single unified approach, often suffering from negative transfer in pretrain-finetune pipelines. This work introduces a Decoupled and Divergence-Conditioned Prompt design for Dynamic Graph Foundation Models to separate domain-specific knowledge and condition prompts on domain divergence, improving cross-domain generalization. The results demonstrate better transfer across domains with inconsistent semantic and temporal patterns.
Hypergraph neural networks' expressivity hinges on detecting and counting small higher-order patterns. The authors formalize this using homomorphism densities and establish a strict expressivity hierarchy called WidthWall, linking the patterns an HGNN can capture to its representational power. This framework clarifies fundamental limits of HGNN architectures and guides design choices.
GAAMA proposes a graph-augmented associative memory for agents to maintain persistent long-term, structured memory across sessions. It constructs a concept-mediated knowledge graph via a three-stage pipeline to organize experiences, retrieve relevant memories, and update the graph, enabling coherent and personalized agent behavior. This addresses limits of flat RAG and ego-graph approaches in maintaining relational memory.
T-TExTS presents a KG-based recommendation system to help high school English teachers assemble diverse, thematically aligned texts. Building a KNARM ontology, it targets pedagogical merit rather than surface metadata to expand teaching texts, and demonstrates domain-specific ontology instantiation for data-driven recommendations. The approach supports teacher scaffolding and planning with knowledge-driven selections.
We present an exact verification method for graph neural networks using incremental constraint solving, enabling sound and complete guarantees against perturbations that modify attributes or the graph structure (including edge additions or deletions). The framework supports common GNN aggregation functions and yields precise robustness certificates for high-stakes applications. It advances beyond approximate verification methods.
RAG-GNN is an end-to-end retrieval-augmented GNN framework that fuses dynamically retrieved biomedical knowledge with graph representations for precision medicine. It uses a retrieval projection, gated fusion, and contrastive alignment to integrate literature-derived knowledge, improving functional clustering and interpretation in a cancer signaling dataset (379 proteins, 3,498 interactions, 14 categories).
Why Neighborhoods Matter argues that citation faithfulness in Agentic GraphRAG should be considered at the trajectory level, accounting for traversal context, graph structure, and visited-but-uncited entities during reasoning. Through controlled ablations, the work shows how neighborhood context shapes the fidelity of citations in generated answers.
AlayaLaser introduces an efficient on-disk index layout and search strategy for large-scale high-dimensional vector similarity search. The authors show that, at high dimensionality, compute becomes the bottleneck rather than I/O, and present layout and algorithmic optimizations to reduce compute and improve throughput for graph-based ANNS on disk.