Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 20 papers for 2026-08-24
This paper investigates how uncertainty is shaped in Graph Neural Networks with Bayesian output layers. They observe that predictive uncertainty decreases as latent representations align with low-variance posterior directions, suggesting that uncertainty reduction may be driven by latent-posterior geometry rather than classical posterior contraction.
This work provides a critical audit of spatiotemporal forecasting benchmark datasets and baselines. It shows that a small set of datasets dominate the field and that baselines span from simple historical averages to classical machine learning methods, with evaluation protocols often inconsistent.
The paper investigates backdoor vulnerabilities of Graph Foundation Models on text-attributed graphs under graph-language alignment. It argues that existing attacks are largely single-modality and proposes stealthy cross-modal backdoors that exploit alignment, with discussion of defenses.
FlatLand introduces a personalized graph federated learning approach that embeds each client's data into a tailored Lorentz space (hyperbolic geometry) to better capture heterogeneity. This geometry-aware personalization improves performance across diverse graph data.
TH-GNN is a heterogeneous temporal graph neural network designed to detect shilling attacks by modeling cross-modal interactions among graph and text, and by capturing temporal dynamics. It outperforms text-only or graph-only detectors, demonstrating the benefit of integrated heterogeneous temporal structure.
They propose an interpretable brain-graph learning framework that decomposes information into redundancy, uniqueness, and synergy for rs-fMRI-based disease diagnosis. This information decomposition reveals how information is shared across brain regions, improving diagnostic interpretability beyond connectivity strength.
The paper proposes disaster-conditioned kernel-regularized graph attention to utilize spatial context for building damage classification, enabling event-specific neighborhood reasoning. By accounting for different disaster clustering patterns, the method improves robustness and avoids misusing spatial context.
Hessian Interatomic Potentials (HIPs) directly predict Hessians without derivatives or finite differences, enabling efficient Hessian-based modeling for molecular systems. This approach improves scalability and supports Hessian-informed workflows in ML interatomic potentials.
Degree-Mass Message Passing introduces a lightweight GNN to estimate betweenness centrality rankings for both directed and undirected graphs. It achieves scalable performance with competitive accuracy compared to heavier methods.
PROBE-Web is an interactive system to probe evaluation landscapes for knowledge graph completion models. It lets users compare models under two perspectives—predictive sharpness and robustness to popularity bias—via a user-friendly GUI.
Deterministic and probabilistic neural surrogates for global hybrid-Vlasov simulations learn the spatiotemporal evolution of electromagnetic fields and ion-velocity moments from a few runs, providing both deterministic emulators and probabilistic ones to quantify uncertainty.
FeLoG presents scalable distributed graph embedding with a feedback loop mechanism, where sampling is guided by evolving embedding quality, reducing redundant exploration and boosting embedding efficiency on large-scale graphs.
Continuous-Time Quantum Walks based Graph Neural Network introduces CTQW-based message passing to address over-smoothing and heterophily in graphs, leveraging quantum-walk dynamics for richer representations.
Toward Auto-Research argues that current LLM-based ideation systems treat papers as flat objects, and proposes using category theory to structure paper knowledge graphs to enable generating falsifiable research ideas.
Knowledge-Graph-Gated Defactualization (DSR) provides a framework to separate verifiable facts from stylistic content when generating agentic AI responses, enabling style control without factual leakage while preserving factual correctness.
LingShu is a large symptom-centric contextualized knowledge graph bridging Traditional Chinese Medicine and modern biomedicine, using symptoms as a shared phenotypic layer to connect clinical manifestations to diseases and molecular mechanisms.
Denoising the Future proposes FreqDiff, a frequency-aware diffusion framework for temporal knowledge graph extrapolation, aiming to better separate query-specific evidence from historical noise and improve future fact prediction.
Adapting Knowledge Graphs for Behavior Denoising leverages knowledge graphs to provide evidence to distinguish persistent preferences from temporary exploration in sequential recommendations, improving robustness by denoising behavior signals.
Complete Cyclic Subtask Graphs study cyclic subtask graphs for tool-using LLM agents where subtasks are fully connected, evaluating Spec-Cyc and Gen-Cyc on multiple benchmarks and analyzing flexibility, cost, and bottlenecks in long-horizon workflows.
Columnar-Embedder proposes a biologically inspired cortical architecture for binary sparse distributed graph representations, enabling efficient and generalizable graph embeddings with binary sparse codes, contrasted with transductive, gradient-based methods.