Showing 19 papers for 2026-04-10
BiScale-GTR introduces a fragment-aware multi-scale graph transformer for molecular representation learning. It integrates fragment-level and scale-aware biases to capture patterns across multiple molecular scales, addressing the dominance of local message passing in hybrid GNN-Transformers. Experiments show improved molecular property prediction and more versatile representations.
Toward a universal foundation model for graph-structured data proposes a reusable foundation model for graph analysis. Graphs are central in biomedical research and existing GNNs are typically trained on a single dataset, limiting generalization. The work outlines a scalable approach—potentially large-scale pretraining and multitask objectives—that enables reusable graph representations across diverse graphs and tasks.
From Load Tests to Live Streams presents a graph embedding-based anomaly detection system for microservice architectures. Built on a GCN-GAE, it learns unsupervised node-level representations from directed weighted graphs to identify under-represented or anomalous services. The approach helps bridge gaps between synthetic load tests and real event traffic.
GraphWalker proposes a graph-guided in-context learning framework for clinical reasoning on electronic health records. It addresses perspective limitation and cohort-awareness by leveraging graph structures to guide exemplar retrieval and reasoning in LLMs, improving clinical reasoning accuracy on EHR data.
Graph Neural ODE Digital Twins for Control-Oriented Reactor Thermal-Hydraulic Forecasting Under Partial Observability introduces a physics-informed GNN-ODE digital twin to forecast plant-wide thermal-hydraulic states with partial observability. It combines message-passing GNNs with neural ODEs to deliver accurate, real-time forecasts suitable for control applications.
DynLP introduces a parallel dynamic batch update for label propagation in semi-supervised learning. It enables efficient, incremental updates when new data batches arrive, avoiding full recomputation and speeding up semi-supervised inference on graphs.
BadImplant presents a multi-target backdoor attack for graph classification. By injecting multiple triggers, the attacker can redirect predictions to different target labels, revealing vulnerabilities in graph neural networks and motivating defense strategies.
k-Maximum Inner Product Attention for Graph Transformers proposes a scalable attention mechanism to reduce quadratic memory and computation in graph transformers while preserving expressive power. The approach enhances scalability to large graphs and is analyzed in the context of GraphGPS.
Resource-constrained Amazons chess decision framework integrating large language models and graph attention proposes a lightweight hybrid framework for the Game of the Amazons. It enables weak-to-strong generalization under resource constraints by combining LLM reasoning with graph attention.
Knowledge Graphs Generation from Cultural Heritage Texts combines LLMs and ontological engineering in a five-step ATR4CH framework to extract and generate knowledge graphs from cultural heritage documents. A case study on authenticity debates validates the approach by integrating annotation models, ontologies, and LLM-based extraction.
A Comparative Study of Demonstration Selection for Practical Large Language Models-based Next POI Prediction analyzes how different demonstration strategies (e.g., random, embedding-based) affect the accuracy of next point-of-interest predictions with LLMs, offering practical guidelines for in-context learning in POI tasks.
Ontology-based knowledge graph infrastructure for interoperable atomistic simulation data presents an ontology-driven knowledge graph framework to represent and integrate atomistic simulation data. It combines domain ontologies with a software framework to capture data from datasets and simulation workflows, enabling standardized metadata and provenance.
"Don't Be Afraid, Just Learn": Insights from Industry Practitioners to Prepare Software Engineers in the Age of Generative AI surveys 51 industry practitioners and 11 follow-up interviews to understand hiring practices, required skills, curricular gaps, and how universities should adapt to the GenAI era.
A Graph-Enhanced Defense Framework for Explainable Fake News Detection with LLM proposes a graph-enhanced detection framework that uses LLMs with externally retrieved reports as evidence. It emphasizes the need for reliable evidence and aims to provide explanations across all aspects of a claim.
ReRec: Reasoning-Augmented LLM-based Recommendation Assistant via Reinforcement Fine-tuning introduces a reinforcement fine-tuning framework to improve LLM reasoning in complex recommendation tasks. It features components such as dual-graph enhanced signals to guide reasoning and improve recommendation quality.
Task-Adaptive Retrieval over Agentic Multi-Modal Web Histories via Learned Graph Memory introduces ACGM, a learned graph-memory retriever that builds task-adaptive relevance graphs over multi-modal agent histories. It optimizes relevance graphs with policy-gradient methods to improve retrieval for downstream tasks.
Rethinking Recommendation Paradigms: From Pipelines to Agentic Recommender Systems argues for moving away from static multi-stage pipelines toward Agentic Recommender Systems where modular components are guided by intelligent agents. This design aims to enable more adaptive, goal-driven recommendations in heterogeneous data environments.
SynQL: A Controllable and Scalable Rule-Based Framework for SQL Workload Synthesis for Performance Benchmarking introduces a controllable, rule-based SQL workload synthesis framework. It combines rules with LLMs to generate realistic workloads while constraining schemas to avoid hallucination and preserve privacy.
OpenGLT: A Comprehensive Benchmark of Graph Neural Networks for Graph-Level Tasks presents a broad benchmark suite for graph-level tasks. It addresses limitations of prior evaluations by expanding dataset coverage, architecture diversity, task scope, and realistic benchmarking scenarios.