Showing 19 papers for 2026-05-23
Ex-GraphRAG replaces the GNN encoder with a Multivariate Graph Neural Additive Network (M-GNAN) to make evidence routing in Graph-Augmented LLMs interpretable. It addresses the inability to audit node contributions caused by iterative neighborhood aggregation and enables faithful auditing of which retrieved entities influenced the model output.
We show that higher-order subgraph attribution via GNN-LRP can be computed efficiently with a new message-passing approach that avoids summing exponentially many walks. The method significantly reduces computational complexity, making practical, higher-order explanations feasible.
We study three graph tokenizations for Transformers—spectral, random-walk, and adjacency—and show that tokenization fundamentally constrains transformer expressivity. Different tokenizations induce distinct depth regimes and performance trade-offs, underscoring the importance of tokenization design for graph-aware transformers.
Mini-batch training of GNNs induces implicit regularization through subgraph sampling. The simplest scheme, Random Node Sampling (RNS) on the induced subgraph, matches or even beats full-graph training on most datasets while saving substantial training time.
We evaluate deep ensembles for message-passing GNNs and find surprisingly limited gains in uncertainty quantification. Across seven datasets, ensembles provide little improvement over a strong single model, with any gains mainly arising from better single-model performance or calibration.
We systematically compare four explanation methods (SA, Integrated Gradients, GNNExplainer, and Layer-wise Relevance Propagation) for disease-relevant structure in a breast cancer RNA-seq projected PPI network. Using synthetic benchmarks with ground-truth, we show that explanations reveal a topological signature of disease-associated hubs, supporting biologically meaningful interpretability.
A2QTGN is a hybrid quantum-classical framework that couples adaptive amplitude encoding with a Temporal Graph Network backbone for dynamic link prediction. By integrating quantum-inspired encoding with a scalable GNN, it aims to better capture concurrent and rapidly changing interactions in large dynamic graphs.
SPECTRA is a spectral-domain-aware graph generation method designed to improve regression for underrepresented molecular property values. It uses a rarity-aware budgeting scheme to focus generation where data is scarce, boosting predictive accuracy in highly relevant but under-sampled ranges.
Billion-Scale Graph Foundation Models (GraphBFF) present a practical recipe for building billion-parameter GFMs on large-scale heterogeneous graphs. The work outlines a GraphBFF Transformer and end-to-end pretraining and adaptation strategies to enable scalable, real-world graph foundation models.
Geometry-Induced Diffusion on Graphs introduces mu-ChebNet, a spectral GNN that learns a node-wise weight function mu before applying Chebyshev filters. The learned mu modifies the propagation geometry via a weighted Laplacian without changing the graph topology, improving long-range information diffusion.
MapTab is a multimodal benchmark to assess MLLMs on multi-criteria route planning in heterogeneous graphs. It requires models to ground map visuals and route attributes (e.g., time, cost) and perform holistic reasoning under multiple criteria.
Beyond One-Size-Fits-All proposes adaptive subgraph denoising for zero-shot graph learning with LLMs. By denoising subgraphs to improve cross-modal alignment between graphs and text, it enhances zero-shot predictive performance.
MemReward presents a graph-based experience memory for LLM reward prediction under limited labels. The memory stores past rollout experiences to help predict rewards when ground-truth labels are scarce, reducing annotation burden.
Position argues that graph condensation should move beyond full-dataset training and model-dependent methods. The paper advocates resetting and developing model-agnostic condensation strategies to improve scalability and robustness for large graphs.
Format-Constraint Coupling shows that on statistical tables, the interaction between serialization format and schema constraints can affect knowledge graph fidelity in a super-additive way. The results demonstrate that applying a schema to a mismatched format can trigger substantial degradation across datasets.
Knowledge Graph Re-engineering Along the Ontological Continuum advocates re-engineering KG representations along an ontological continuum to support neuro-symbolic AI. It discusses aligning vocabularies, axioms, and ontologies and leveraging GenAI automation to improve KG interoperability and reuse.
Dynamic Hypergraph Representation Learning builds dynamic hypergraphs from multivariate time series without prior structure. The model learns hypergraph representations that capture higher-order relationships across variables as time evolves.
Case-Aware Medical Image Classification uses multimodal knowledge graphs and reliability-guided refinement to enable case-aware reasoning for explainable medical image diagnosis. Given an image, the method integrates historical similar cases and symptoms to support diagnosis with reliability-aware refinements.
CatalyticMLLM proposes a Graph-Text Multimodal LLM for catalytic materials, unifying graph-based representations with text modalities to address inconsistencies between generation and property-prediction spaces. This integration aims to improve property prediction and enable more effective generation-driven inverse design.