Showing 31 papers for 2026-07-28
This paper develops a graph attention network (GAT) framework to predict soil microplastics and organic matter by modeling spatial dependencies among 91 georeferenced samples. By fusing spatial coordinates, soil properties, and land-use data in a two-layer GAT, it captures local interactions and delivers strong predictive performance (RMSE 625.06 and R^2 0.87 for microplastics), with promising results also reported for organic matter.
To improve anomaly detection in cyber-physical systems with scarce labeled anomalies and limited normal data, this work identifies the limitations of existing graph-based models that can learn spurious correlations and unstable topologies in small-scale settings. It introduces DPR-GM (Domain-Prior-Regularized Graph Model), which injects domain priors as regularization to guide learning and enhance robustness.
ParasGB provides a public graph benchmark suite for parasitic estimation in AMS (analog/mixed-signal) circuits, addressing the lack of high-fidelity RC benchmarks for reproducible evaluation. It offers high-quality RC datasets and evaluation protocols to support reproducible, parasitics-aware modeling with graph neural networks during early-stage design exploration.
MEGA-CL is a foundation graph neural network framework for universal molecular ADMET prediction. It combines self-supervised contrastive learning with a multi-head external attention mechanism and an enhanced message passing architecture to jointly model local chemical substructures and global inter-graph relationships while mitigating over-smoothing.
We propose a decentralized GNN-based multi-agent control strategy to mitigate traffic shockwaves in sparse Vehicular Ad-hoc Networks (VANETs). The method enables cooperative control among vehicles without relying on global traffic state information, addressing practical deployment in early-stage CAV networks.
This survey reviews Graph Transformers (GTs), covering architectures, theoretical foundations, and applications. It discusses how GTs address issues such as over-smoothing and over-squashing in graph learning and provides a comprehensive synthesis of recent progress and challenges.
We propose CHARM, a method to detect LLM hallucinations by unifying signals into attributed graphs: tokens become nodes, edges follow attentional flows, and both carry features from attention scores and activations. Graph neural networks are then applied to detect hallucinations as a graph-learning task.
Plain Transformers are shown to be surprisingly powerful link predictors, achieving competitive performance for link prediction without relying on explicit graph-structural heuristics or heavy node embeddings. The study argues for the effectiveness and scalability of plain Transformer architectures in graph tasks.
This work uses undirected Cayley graphs to represent finite groups and trains a graph neural network to classify groups by solvability, without relying on explicit algebraic features. The experiments on 200 groups (120 solvable, 80 non-solvable) probe how well GNNs can learn algebraic properties from graph structure.
The authors introduce topology-aligned inductive biases for molecular property prediction, where atoms map to fixed computational units and bonds govern interacting pairs. They instantiate two architectures, Iso-QGNN (a quantum circuit) and Iso-CGNN (a classical message-passing network), both aligned to the molecular bond graph to improve data efficiency.
INSIGHT is a graph neural network that predicts survival directly from routine histology images, producing patient-level spatial risk scores. Trained on TCGA and SURGEN and validated on independent cohorts, it shows superior prognostic performance compared with pTNM staging and reveals spatial epithelial-immune prognostic axes.
FB-GNN-MBE integrates a fragment-based graph neural network into many-body expansion theory to efficiently reproduce first-principles potential energy surfaces for large systems, with data-adaptive transfer learning enabling cross-system PES accuracy.
Spectral Flow Certificates (SFCs) are scalar certificates computed from a graph's normalized Laplacian that predict whether long-range information can propagate on the graph before training, enabling practitioners to assess graph suitability without training a model.
SCOPE and SCION present a benchmark and auditable pipeline for schema induction and fusion from text. SCOPE is a train-text-only corpus-to-schema benchmark built from 24 information extraction sources, producing gold schema graphs for evaluation, while SCION provides an auditable reference pipeline for schema induction.
FrED proposes a probabilistic framework for external data influence attribution that operates in a black-box setting. It fuses continuous feature similarities with discrete, domain-specific knowledge graphs to provide grounded, interpretable attribution for external data influence.
Graph-theoretic fragmentation with covariant direct molecular force learning enables coupled-cluster accuracy AIMD for fluxional systems by fragmenting molecules and learning nuclear forces with covariance-aware ML, avoiding problematic automatic differentiation on energy surfaces and improving Jacobian handling for link atoms.
MA-DAR introduces Manifold-Aligned Dynamic Adaptive Routing for continual temporal knowledge graph reasoning, addressing how replayed representations are integrated with current ones to avoid norm domination and semantic drift during continual learning on TKGs.
MineValiCoder advances reliable code generation with test case quality mining and bipartite-graph mutual validation, reducing the risk of misleading feedback from faulty tests and resolving conflicts from mixed-quality tests in LLM-driven TDD.
MedKGent presents a large-language-model agent framework to construct temporally evolving medical knowledge graphs from 10 million PubMed abstracts (1975–2023), using two specialized agents to incrementally build the KG daily and maintain up-to-date medical knowledge.
This study analyzes short-term-to-long-term memory transfer in reinforcement learning under partial observability. It casts memory management as a neuro-symbolic decision problem, using per-item Q-learning to decide for each observed triple whether to keep or drop it before long-term storage, enabling scalable memory with variable-sized buffers.
RankGraph-2 presents an end-to-end lifecycle co-design for billion-node graph-based recommendation, jointly optimizing graph construction, representation learning, and real-time serving. It introduces a co-trained cluster index to enable fast online retrieval and reduces costly KNN at serving time by integrating index learning into training.
Melo is a production LLM-powered music recommendation agent deployed at NetEase Cloud Music. It is built as a deterministic five-node state graph over heterogeneous tools, guided by prompts and a state-machine rather than fine-tuned controllers. The authors emphasize robustness: production bottlenecks are failure detection and recovery, addressing entity hallucination and other failure modes.
The paper addresses reliable interpretation of historical documents in knowledge graphs using agentic retrieval with query inference and execution. It aims to balance the generalization ability of LLMs with accountability required by archival institutions, providing mechanisms for traceable reasoning and verifiable outputs.
Structure Over Scale introduces HCG-RAG, a schema-constrained causal-graph retrieval-augmented generation framework. Instead of exhaustive extraction, it distills a corpus into a fixed, typed vocabulary of causal variables and builds a two-tier graph that supports scalable, reasoning-focused retrieval.
We propose PCA-GAT, a graph attention model that integrates factual industrial knowledge and normative decision constraints to recommend machining process plans. By encoding material properties, features, and quality requirements, PCA-GAT unifies retrieval and ranking with constraint awareness, enabling more reliable and compliant process plans.
This survey reviews how graph structures are used to augment context for re-ranking in information retrieval and related tasks. It discusses graph representations, relational reasoning, and practical challenges such as scalability, dynamic updates, and integration with RAG/QA systems, and outlines opportunities for future research.
TextBridgeGNN proposes a text-guided pre-training framework to bridge cross-domain recommendations. It addresses two key challenges: non-transferable ID embeddings across domains and structural incompatibility between heterogeneous domain graphs. By leveraging textual information to guide cross-domain alignment, it enables better transfer and recommendation performance.
This work presents a controlled, tool-mediated agentic GraphRAG architecture for auditable natural-language analysis of public commercial registries. It converts publications from the Swiss Official Gazette of Commerce into a Neo4j knowledge graph with millions of nodes, enabling structured reasoning and traceable NL analysis.
CleANN enables efficient full dynamism for graph-based ANNS, supporting insertions, deletions, and queries with maintained indexing efficiency and accuracy. It introduces data structures and update strategies to keep the graph index up-to-date without sacrificing query performance.
This paper presents a simple and fast graph-based filtered ANNS algorithm that incorporates attribute filters (e.g., category, color, brand) into the search process. The method balances retrieval quality and efficiency by integrating filtering directly into graph traversal.
TGMS is a bi-temporal property graph management system that exposes 13 verified temporal operators as agent tools. LLM agents plan operator calls and generate final answers while the system executes the graph evolution and corrections, enabling reliable bi-temporal querying.