Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 13 papers for 2026-08-13
We propose a minimal table-to-graph abstraction where each row is a node, without learning any graph structure. A message-passing GNN operating under the 1-WL limitation is applied to this abstraction, so relational reasoning reduces to a partition of rows. The approach avoids training a graph model or designing explicit schemas, and provides a criterion that does not require graph supervision.
We study model extraction attacks on GNNs in MLaaS settings and show that defenses borrowed from Euclidean-domain methods (e.g., random noise) perform poorly on graphs because of complex topologies. We propose a defense based on model reprogramming to impede extraction while preserving task utility, highlighting the role of topology-aware strategies. The results illustrate trade-offs between security and utility.
We rethink graph counterfactual explanations using discrete diffusion inversion. The discrete search space makes finding valid, minimal structural modifications difficult, and standard methods struggle to produce feasible explanations. Diffusion inversion provides a principled way to generate faithful, sufficient, and understandable graph counterfactuals.
We propose Variational Mixture of Graph Neural Experts (VMoGE) for Alzheimer's disease recognition across frequency bands in EEG brain networks. The framework combines multi-band EEG representations with variational GNNs and a mixture-of-experts to model heterogeneous spectral patterns. It improves discrimination between Alzheimer's disease and frontotemporal dementia.
We show that existing spectral-energy-based spatio-temporal graph anomaly detection methods mostly detect anomalies via increased spectral variation, missing camouflaged anomalies that decrease variation. We propose a node-level spectral energy formulation that is fully compatible with message passing and can detect both increases and decreases in energy. Experiments on multiple datasets demonstrate the approach's effectiveness.
We address credit risk detection at WeChat Pay scale, where billions of users are connected by heterogeneous risk graphs. Graph neural networks capture complex dependencies but face scalability bottlenecks. We propose a scalable GNN framework designed for billion-scale data to improve fraud detection while maintaining efficiency.
The paper introduces Causal-Guided Representation Learning (CGRL) to improve node-level out-of-distribution generalization in GNNs. It argues that standard methods pick up environment-specific noise and spurious correlations, hurting robustness. CGRL explicitly models the fine-grained latent geometry with causal guidance to enhance robustness.
We demonstrate an autonomous LLM research loop that optimizes crystal graph networks for band-gap prediction on MatBench. The approach uses a general-purpose LLM to guide architectural search and hyperparameter tuning, building strong models without external pretraining. The results illustrate the potential of autonomous language-model-driven design in materials science.
BEST-KAG proposes a multimodal knowledge-driven framework for building engineering standards question answering. It enables multi-clause reasoning and multimodal knowledge utilization, with clause-level evidence linkage to support answers. This framework moves beyond keyword-based retrieval for standards QA.
We propose Graph-Structured Rubrics, which compile rubrics into a typed evaluation graph before observing responses. Criterion nodes elicit judgments; transformation, reduction, and gating operators compose them through named ports; and a task-specific Readout maps the unique sink to a final score. The approach supports structured, explainable evaluation by LLM judges.
We present KG-DML, a framework for constructing Dynamic Master Logic Models as Knowledge Graphs for complex system diagnostics, leveraging Retrieval-Augmented Generation and LLMs. The framework automates generation or augmentation of the knowledge graph from system descriptions. This reduces reliance on expert interpretation and yields scalable, queryable DML representations.
We propose HSTGFormer, a Hyper Spatial-Temporal Graph Transformer for 3D human pose estimation. It reformulates spatial-temporal reasoning as localized coupled graph aggregation over joint-time nodes. The graph-enhanced Transformer models the local spatio-temporal structure to improve pose estimation accuracy.
We introduce GALLM, Graph-Aware Large Language Models for sequential recommendation. By incorporating collaborative signals encoded in user-item graphs, the framework complements the LLM backbone with graph-based representations. The method aims to capture global collaboration patterns beyond within-sequence dependencies.