Showing 23 papers for 2026-04-09
BiScale-GTR introduces a fragment-aware, multi-scale graph transformer for molecular representation. It combines fragment-level information with multi-scale graph structures to capture patterns across different molecular scales, reducing reliance on local message passing and uniting GNN inductive biases with Transformer global receptive fields. The approach yields improved molecular property prediction performance.
Toward a universal foundation model for graph-structured data advocates the idea of a reusable foundation model for graphs. It notes that graphs are pervasive in biomedicine and beyond, but current GNNs are trained on single datasets and learn representations tied to specific node features, topologies, and labels. The work outlines a framework for a universal graph foundation model that can generalize across graphs, features, topologies, and tasks.
From Load Tests to Live Streams proposes a graph embedding-based anomaly detection system for microservice architectures. It uses an unsupervised node-level embedding approach (GCN-GAE) to learn structural representations from directed, weighted service graphs, enabling detection of under-represented or anomalous services during live traffic streams.
GraphWalker presents a Graph-Guided In-Context Learning framework for clinical reasoning on electronic health records. It addresses challenges of perspective alignment, cohort awareness, and demonstration selection for in-context learning, aiming to improve the reliability and relevance of LLM-based clinical reasoning. The method uses graph signals from EHRs to guide demonstrations and reasoning.
Graph Neural ODE Digital Twins propose a physics-informed GNN-ODE framework for control-oriented reactor thermal-hydraulic forecasting under partial observability. It combines graph-based neural dynamics with neural ordinary differential equations to provide accurate, fast, and robust plant-wide state predictions when sensor data are incomplete, enabling real-time supervisory control.
DynLP introduces parallel dynamic batch updates for label propagation in semi-supervised learning. It enables incremental updates when new data batches arrive, avoiding full re-computation and improving efficiency and scalability of label propagation in streaming or iterative settings.
BadImplant delivers a injection-based multi-targeted backdoor attack for graph classification. It implants multiple triggers into a GNN to redirect predictions to different targets, enabling coordinated backdoors across several classes. The work analyzes attack effectiveness and discusses potential defenses.
k-Maximum Inner Product Attention for Graph Transformers proposes a scalable attention mechanism to balance efficiency and expressiveness on large graphs. By replacing quadratic all-to-all attention with a k-maximum inner product scheme, it improves scalability, with analyses on the expressive power of GraphGPS.
Resource-constrained Amazons chess decision framework integrates large language models and graph attention to enable lightweight, robust play under limited compute. The framework targets weak-to-strong generalization in decision making by combining LLM reasoning with graph-based guidance.
EAGLE introduces Edge-Aware Graph Learning for proactive delivery delay prediction in smart logistics networks. It blends graph-structured insights with edge-level information to forecast delays, bridging tabular/time-series methods and networked data to capture spatial-temporal dependencies.
OntoTKGE proposes ontology-enhanced temporal knowledge graph extrapolation. It presents an encoder-decoder framework that leverages ontological knowledge to alleviate sparsity in historical interactions, enabling entities with few past events to inherit behavioral patterns from same-concept entities.
Generative AI for Video Trailer Synthesis describes a shift from extractive heuristics to autoregressive creativity, leveraging LLMs, multimodal models, and diffusion-based video synthesis to construct coherent, representative trailers around key moments.
Phase-Associative Memory introduces PAM, a complex-valued recurrent sequence model with associative memory in complex space. It stores state as a complex matrix updated via outer products and retrieves via conjugate inner product, achieving competitive perplexity with ~100M parameters on WikiText-103 while incurring higher arithmetic costs.
LanG: governance-aware agentic AI platform for unified security operations. It offers open-source tooling for unified security operations, including a Unified Incident Context Record with a correlation engine and an Agentic AI Orchestrator to coordinate actions and improve SOC workflows.
A Formal Security Framework for MCP-Based AI Agents articulates a threat taxonomy, verification models, and defense mechanisms for MCP-based AI agents. It provides a unified approach to characterizing, validating, and mitigating risks when LLM agents use external tools and data sources.
LaSM proposes a Layer-wise Scaling Mechanism to defend GUI agents against pop-up environmental injection attacks. By applying defenses at multiple layers, it achieves robust protection with moderate retraining costs.
Knowledge Reasoning Language Model unifies knowledge and language for inductive knowledge graph reasoning. It blends KG structure with language modeling to improve inductive reasoning on open-domain KGs, leveraging LLMs and KGFMs to handle component uncertainty.
Agora teaches consensus-finding with AI personas grounded in human voice. It uses an AI-powered platform to organize authentic human perspectives on policy issues, helping users practice consensus-building through structured dialogue and synthesis.
UnWeaving the knots of GraphRAG finds that VectorRAG is almost enough for many retrieval-augmented generation tasks. It analyzes the relative benefits of graph-based vs vector-based retrieval for multi-hop questions, suggesting that vectors often suffice while graphs offer niche gains.
Beyond Case Law evaluates structure-aware retrieval and safety in statute-centric legal QA with SearchFireSafety. Using fire-safety regulations as a representative domain, the benchmark tests how evidence is distributed across hierarchically linked statutes and how models handle incomplete context.
This paper proposes a persona-driven session-based recommender that addresses anonymity in SBRS by explicitly modeling latent user personas and integrating LLM-generated item representations with a heterogeneous knowledge graph to improve personalization under sparse or cold-start conditions. It leverages LLMs to enrich item representations and uses a heterogeneous knowledge graph to capture user preferences across sessions.
The work investigates what makes an ideal quote, showing that users prefer quotes that are 'unexpected yet rational' within a given context and that novelty is a key criterion. It proposes a recommendation approach that emphasizes novelty while maintaining contextual rationality, addressing the shortfalls of existing models that focus mainly on topical relevance. The study also discusses how to balance surprise with coherence in automated quote suggestions.
This paper presents an ontology-based infrastructure for interoperable atomistic simulation data, enabling knowledge graphs that represent workflows, provenance, and metadata across sources. The approach combines domain ontologies with a software framework to capture data from existing datasets and directly from simulation workflows at generation, normalizing heterogeneous data into a unified knowledge graph.