Showing 10 papers for 2026-04-20
This paper analyzes attribute-level heterophily in graphs and shows that connected nodes can have dissimilar attributes, which undermines homophily-based unsupervised GNNs for anomaly detection. It argues these limitations make existing methods impractical in real graphs and proposes NK-GAD, a neighbor-knowledge-enhanced framework for unsupervised graph anomaly detection that leverages local neighbor information to detect anomalies more reliably.
In the face of failures that can partition a UAV swarm, centralized recovery methods require global topology and become communication-heavy. Decentralized heuristics and MARL struggle to scale with swarm size and damage. The authors propose a physics-informed graph adversarial imitation learning framework enabling zero-shot, scalable resilience in decentralized UAV swarms by leveraging graph structure and physics constraints.
Current GNNs mainly capture pairwise relations, but real systems contain higher-order topologies. We introduce a structure-aware simplicial spatiotemporal neural network that uses simplicial complexes to model higher-order relationships in space-time data, enabling more accurate and scalable modeling of complex networks.
Hyperbolic conservation laws are challenging for neural surrogates due to shocks and complex waves. We present a structure-preserving graph neural solver that respects intrinsic PDE structures, providing physically admissible, robust surrogates for parametric studies and many-query tasks, while improving efficiency over classical solvers.
This survey examines how graphs are integrated with LLMs to enhance reasoning, retrieval, and generation. It categorizes design choices and applications, clarifies when graph-LM integrations are beneficial, and surveys methods for reasoning, knowledge retrieval, and guided generation.
We propose using a knowledge graph to store domain data, ML results, and explanations, linking them to improve interpretability in manufacturing ML. A selective retrieval method surfaces relevant triplets to users, enabling explanations that align with domain concepts and facilitating user understanding.
We present an exascale workflow for materials discovery using atomistic graph foundation models built on HydraGNN, jointly trained on 16 datasets (544+ million structures across 85+ elements) with a multi-task head design and a scalable data pipeline. On the Frontier supercomputer, we run multiple DeepHyper hyperparameter campaigns and promote top models to sustained 2,048-node training, yielding a PaiNN-based lead model.
Persona-guided prompting reveals motivated reasoning in LLMs, showing human-like biases that can influence judgments and polarization. The paper investigates how assigning a persona affects LLM reasoning and whether such conditioning leads to biased or biased-like outputs.
We examine analytic flexibility in using LLMs to simulate human data, showing that choices like model selection, prompts, sampling, and demographic context can materially affect how silicon samples resemble human data. Across two studies, we demonstrate how configuration choices alter external validity.
We propose a preference-aligned user simulator for recommendations that leverages user feedback to calibrate simulated preferences, addressing ambiguity and noise in feedback and enabling more realistic evaluation and optimization of RSs.