Showing 16 papers for 2026-06-11
We propose probabilistic contrastive pretraining for multi-task ADME property prediction. The method uses 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 from latent codes, and leverages a contrastive objective to improve multi-task ADME predictions under noisy and data-limited conditions.
GraphInfer-Bench introduces a benchmark to test whether LLMs can perform graph inference. It asks models to produce open-ended answers that rely on the graph structure in a way no single node or path can fully support. The benchmark highlights limitations of existing graph-QA protocols and includes tasks requiring algorithmic reasoning, node classification, and multi-hop inference.
Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teaching. We address few-shot learning on text-attributed graphs by coupling GNNs and LLMs in a co-teaching framework, mitigating the respective weaknesses: GNNs struggle with cold nodes, LLMs with text-ambiguous nodes. By moving beyond a single golden teacher, the two models supervise each other to improve TAG performance.
TAROT: Task-Adaptive Refinement of LLM-prior Graphs for Few-shot Tabular Learning. We propose TAROT, a framework that adaptively refines LLM-derived graph priors to suit each tabular task in few-shot settings. This reduces privacy risks and computational overhead compared to directly prompting raw tabular data to LLMs, while achieving task-specific improvements.
Probabilistic Salary Prediction with Graph Attention Networks and a Mixture Density Network. The model, GAT-MDN, uses graph attention to capture relational effects among attributes like location, occupation, and industry, and outputs a probabilistic salary distribution via a Mixture Density Network. This addresses uncertainty and multi-modality in real-world salary data.
GraspLLM: Towards Zero-Shot Generalization on Text-Attributed Graphs with LLMs. We propose a framework that enables zero-shot generalization across TAGs by leveraging the semantic strengths of LLMs together with graph-aware representations to transfer knowledge across graphs and tasks.
Neuro-Relational Programs: Unifying Queries and Neural Computation over Structured Data. We introduce Neuro-Relational Programs (NRPs), a declarative query language for relational databases whose facts carry numeric vector embeddings. NRPs extend Datalog with neural embeddings to jointly process relational content and vector facts, enabling seamless integration of symbolic queries and neural computation.
GILT: An LLM-Free, Tuning-Free Graph Foundational Model for In-Context Learning. We present GILT, a graph foundational model that operates without LLMs or task-specific tuning and supports in-context learning on graphs. By leveraging self-supervised graph pretraining and a tuning-free design, GILT enables robust, flexible inference across heterogeneous graphs.
SpaTeoGL: Spatiotemporal Graph Learning for Interpretable Seizure Onset Zone Analysis from Intracranial EEG. SpaTeoGL jointly learns window-level spatial graphs among iEEG electrodes and a temporal graph across time windows, with a smooth graph signal processing formulation to produce interpretable seizure networks.
Intention Driven Identification of In-Possession Match Phases in Association Football through Temporal Graph Learning. The framework identifies in-possession phases from spatiotemporal tracking data using a hierarchical phase model guided by tactical intentions. The study analyzes seven German Bundesliga matches (25 Hz TRACAB) and defines three phases based on evolving intentions.
Lung-R1: A Knowledge Graph-Guided LLM for Pulmonary Diagnostic Reasoning. Diagnosing pulmonary diseases requires integrating heterogeneous evidence amid phenotypic variability and cross-disease overlap. Although large language models (LLMs) have shown progress on pulmonary knowledge question answering (QA) and information-processing tasks, reliable pulmonary diagnosis requires patient-specific, relation-aware reasoning over electronic medical record (EMR) evidence rather than isolated knowledge recall. We define this gap between pulmonary knowledge and case-level diagnostic reasoning as the Pulmonary Knowledge-to-Diag
Automating Geometry-Intensive Compliance Checking in BIM: Graph-Based Semantic Reasoning Framework. Automating compliance check for geometry-intensive regulations remains a significant technical bottleneck in Building Information Modeling (BIM). SGR-BIM is proposed as a Spatial-Geometric Reasoning System for BIM that performs graph-based semantic reasoning to automate regulatory compliance checks across geometry-heavy regulations, enabling multi-hop reasoning over multiple building entities.
LLMs+Graphs: Toward Graph-Native, Synergistic AI Systems. Large Language Models (LLMs) have advanced rapidly, but their limitations in structured and multi-hop reasoning underscore the need for graph-native, synergistic AI systems. The paper outlines three complementary synergies: enabling retrieval-augmented LLM reasoning with graphs, graph-based reasoning for structured data, and synergistic execution to ground in graph computation.
MLaGA: Multimodal Large Language and Graph Assistant. We introduce the Multimodal Large Language and Graph Assistant (MLaGA), an approach for graph-based analysis on multimodal graphs where nodes carry diverse attribute types such as text and images, enabling richer integration of linguistic and visual information with graph reasoning.
Graph2Idea: Retrieval-Augmented Scientific Idea Generation with Graph-Structured Contexts. Generating novel ideas benefits from graph-structured contexts that connect papers across problems, methods, mechanisms, and findings. Graph2Idea uses retrieval-augmented generation with graph contexts to make these cross-paper relations explicit and traceable.
U-HNSW: An Efficient Graph-based Solution to ANNS Under Universal Lp Metrics. ANNS-U-Lp is challenging because it requires queries under all p in (0,2] simultaneously. U-HNSW proposes a graph-based index that delivers efficient ANN search for universal Lp metrics, outperforming MLSH (LSH-based) approaches on query speed for fixed p and enabling a single index for all p.