Showing 15 papers for 2026-06-12
We propose a molecular graph-transformer pretraining framework that combines chemistry-specific self-supervision with contrastive mutual information learning (cMIM). It encodes molecular graphs into latent variables, reconstructs SMILES strings from the latent codes, and augments the contrastive objective to improve multi-task ADME property prediction under noisy, interdependent, and data-limited conditions.
Text-attributed graphs (TAGs) pose challenges for few-shot learning. We introduce an LLM-GNN co-teaching framework where large language models and graph neural networks mutually supervise each other, leveraging their complementary strengths to improve learning on scarce labels and cold nodes.
We present GAT-MDN, a unified framework combining Graph Attention Networks with a Mixture Density Network to yield probabilistic salary predictions. The model captures inherent uncertainty and multimodality in pay data, leveraging hierarchical and semantic relationships among locations, occupations, and industries.
GraspLLM enables zero-shot generalization on text-attributed graphs by leveraging LLMs to extract transferable structural patterns across graphs and tasks. The framework addresses cross-graph and cross-task transfer, enhancing robustness and generalization.
Neuro-Relational Programs (NRPs) define a declarative query language for relational databases where facts carry numeric vector embeddings, unifying symbolic querying with neural computation. Building on Datalog, NRPs enable joint reasoning over embeddings and relational content.
GILT is an LLM-free, tuning-free graph foundational model designed for in-context learning on graphs. It handles highly heterogeneous graphs by relying on a small, specialized model and graph-aware strategies to generalize across unseen topologies, feature spaces, and label sets.
SpaTeoGL introduces a spatiotemporal graph learning framework for interpretable seizure onset zone analysis from intracranial EEG. It jointly learns window-level spatial graphs among electrodes and a temporal graph across windows, informed by smooth graph signal processing.
We propose a data-driven framework to identify in-possession match phases in football from spatiotemporal tracking data. The model defines a hierarchical phase structure guided by three tactical intentions to reveal how possession unfolds.
G-Long presents a graph-enhanced memory management approach for long-term dialogue agents. By using a fine-tuned small language model to manage structured memory, it improves memory efficiency and consistency over long conversations.
Transformer-Guided Graph Attention enables direct cardiac mesh reconstruction from medical imagery by integrating transformer-guided attention with graph-based reconstruction, forming a structural digital twin for precision cardiology.
We propose a lightweight, interpretable transformer constructed by unrolling a mixed-graph optimization for traffic forecasting. It uses an undirected spatial graph and a directed temporal graph to predict future signals with an emphasis on interpretability.
KG-ER offers a conceptual schema language for knowledge graphs that abstracts over representations (relational DBs, property graphs, RDF) while capturing the semantics of stored information.
Structuring The Future introduces Spiffy, a speculative decoding method for diffusion LLMs that accelerates inference while provably preserving the model’s output distribution using calibrated draft graphs.
Analog Quantum Asynchronous Event-Based Graph Neural Networks (QA-AEGNNs) map streaming event data onto neutral-atom quantum hardware to implement asynchronous GNNs leveraging Rydberg interactions for efficient processing.
HKVM-RAG introduces a key-value-separated evidence-organization layer for multi-hop RAG. It assembles answer-path hyperedges from cached evidence tuples to support multi-hop reasoning under fixed retrieval budgets.