Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 7 papers for 2026-08-17
We introduce HI-MeshGraphNets, a hierarchical multi-scale GNN framework designed for mesh-based physical simulations. It enables efficient long-range message passing on high-fidelity meshes, mitigating over-smoothing and reducing memory and compute costs while maintaining accuracy.
We present XIGL, an architecture-agnostic human-in-the-loop method to remove shortcut learning in GNNs. By detecting shortcuts via explanations and guiding training to avoid them, XIGL improves robustness and out-of-distribution generalization.
We propose a Structure-Guided Spatiotemporal Attention Graph Neural Network for traffic flow prediction, which incorporates structural priors into the attention mechanism to better capture spatiotemporal dependencies and offer improved transparency over existing deep models used for traffic forecasting.
We introduce a Graph-Based Reinforcement Learning Framework for structured drift diagnosis and recovery in autonomous LLM agents. The framework provides step-level drift detection, risk assessment, and recovery actions in a plug-and-play manner without re-training the base model, by modeling agent workflows as graphs and learning policies over them.
This exposition surveys the over-squashing problem in graph neural networks, outlining what OSQ is, why it arises, and how it differs from over-smoothing. It reviews current methods, benchmarks, and challenges, highlighting open questions and directions for future work.
SheetCompass introduces Hierarchical Relation Graphs for agentic spreadsheet reasoning. It constructs graphs that encode intra- and inter-sheet relations and spatial layouts, preserving boundaries and semantics, enabling LLMs to reason with global spatial context rather than flattening spreadsheets.
We propose LLM-Guided Graph Generation for structure-based local improvement methods. Our automatic pipeline uses prompts to guide an LLM to generate a graph generator that maps any MiniZinc instance to a uniformly weighted graph, enabling structure-based local search across problems in a problem-agnostic way.