Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 28 papers for 2026-08-19
This paper empirically tests whether fusing uncertainty tools into a single pipeline yields calibrated trust in legal case predictions, rather than merely sharper predictions. Using 1,000 real European Court of Human Rights cases and two frontier LLMs, it compares multiple uncertainty-estimation families as per-fact evidence estimators.
BDIP-Net introduces a dual-interaction graph learning framework to predict properties of bilayer materials. It explicitly models both intra-layer bonds and inter-layer van der Waals couplings to enable efficient property prediction without expensive DFT optimization.
The study disentangles whether heterophily or rarity makes graph neural networks fail. By evaluating six GNNs on five datasets with varying levels of homophily, it finds that homophilic nodes tend to be easier to classify than heterophilic ones.
M-LINKX proposes a multiview graph learning approach for EEG based brain disease detection. It segments long EEG recordings into fixed length clips and learns multiple views to better capture temporal and spatial information for distinguishing AD, MCI, and FTD.
GATTA presents Graph Active Learning with Test-Time Augmentation to improve uncertainty estimates in graph based active learning. The method aggregates predictions across augmented views and uses a consistency based filter to select reliable labels.
FinFraudBench offers a heterogeneous graph benchmark that reflects real financial ecosystems, including customers, cards, merchants, categories, and locations. It aims to align benchmarks with industry realities to better evaluate graph based financial fraud detectors.
Amortised Post-Hoc Explanation with Exact Preservation for Dynamic Graph Anomaly Detectors introduces X-StrGNN, a post-hoc explanation layer that wraps a frozen StrGNN to provide attributions for each flagged edge. The approach amortizes explanation computation while preserving exactness.
NICE proposes scale stable perturbations for GNN explanations by introducing noise corruption perturbations to replace the traditional perturb query approach. It mitigates distribution shifts caused by perturbations and improves the reliability of explanations.
The work unifies graph neural networks through a common layer equation with seven components: update domain, channel set, propagation bank, per-channel message maps, channel-fusion operator, ego/residual map, and update map. The factorization clarifies where information moves versus what moves, revealing shared computations across GNNs.
MRieHy introduces a multi feature Riemannian hypergraph for online test time adaptation of motor imagery brain computer interface decoding. By leveraging hypergraphs built from multiple EEG features and Riemannian geometry, it improves cross day transfer and online adaptation.
The paper argues that the absolute dimension of ker L is not a reliable measure of anti oversmoothing in Sheaf Neural Networks. It proposes an index theoretic criterion to quantify oversmoothing resistance given a sheaf Laplacian, providing a more faithful diagnostic.
Using data from about 11,860 participants in the ABCD study, the paper compares cross sectional, longitudinal, and graph augmented approaches for predicting adolescent substance use onset (alcohol sipping, alcohol use, marijuana use). It reports that graph based methods can offer added predictive value beyond traditional baselines.
BRAID proposes a Best-Response Amortized Iterative Dynamics model that uses a weight tied iterative GNN to learn a direct map from IDS game parameters to Nash equilibrium effort profiles. This replaces costly best response iterations with a single learned forward pass.
The paper introduces domain agnostic neural topic modeling with contextual token level semantic graph representation. By building contextual graphs at the token level, it reshapes embedding geometry to improve topic interpretability across domain specific corpora.
POI Recommendation with LLM-Augmented Multi-Graph Learning and Contrastive Alignment extends LightGCN with two auxiliary item item graphs a semantic graph derived from item descriptions and a spatial graph reflecting geography. A contrastive alignment objective encourages cross graph consistency, addressing cold start and boosting recommendations.
The paper proposes Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing. It employs regime aware routing among multiple experts to identify distinct anomaly mechanisms across heterogeneous networks, enabling targeted intervention.
Not All Neighbors Matter investigates how graph sparsification affects GNN pipelines. Through systematic experiments, it shows that selective edge removal can reduce computation with limited loss in accuracy, and it offers guidelines on which neighbors to retain.
Learning Optimal Dynamic Matching via Graph Neural Networks presents a value based reinforcement learning framework for dynamic matching on evolving weighted graphs. It proves an event time reduction the planner can act immediately after each exogenous event, simplifying the decision process.
EquiPocket introduces an E3 equivariant geometric graph neural network for ligand binding site prediction. By respecting 3D geometry it overcomes rotation sensitivity and irregular protein structures that hinder voxel based 3D CNNs.
Transferable FB-GNN-MBE Framework for Potential Energy Surfaces develops a fragment based GNN integrated with many body expansion to reproduce first principles PES. It demonstrates data adaptive transfer learning and transferability across chemical systems.
The paper argues that knowledge graph reasoning with large language models suffers from a representation mismatch between KG structural context and the model's parametric knowledge. It proposes a Structure-Internalized Rule Language Model that absorbs structural cues into the reasoning process to improve faithfulness in KGR.
This work targets node classification explanations in simplified graph convolutions (SGCs). While methods like LogicXGNN derive global logical rules from explanatory subgraphs, these subgraphs can be redundant and idiosyncratic to individual nodes. The authors propose a logic-based framework to extract global, reusable rules with reduced redundancy, enhancing generalization and interpretability.
The paper tackles RLVR scheduling by estimating sample difficulty online using graph-structured signals. Current strategies either treat samples uniformly or use curricula or other nonuniform allocations; the proposed approach assigns exploration budgets according to estimated difficulty. This improves efficiency by reducing waste on easy samples and ensuring hard but learnable ones receive adequate exploration.
G-ReAct introduces graph-guided deep search where structure and state co-evolve during multi-hop reasoning. This helps preserve intermediate states and constraints across long horizons, mitigating context forgetting and search drift that hinder linear sequential reasoning in LLMs.
VDGR-RAG presents a unified retrieval-augmented generation framework for enterprise knowledge by integrating vectors, directories, graphs, and reflective reasoning. The approach enables better domain routing and exploitation of hierarchical document structures to support robust enterprise QA.
The paper develops a method to guide graph generation for structure-based local improvement by prompting an LLM with semantic guidelines to build a problem-type graph generator. The generator maps any problem instance (in MiniZinc-formatted problems) to a uniform weighted graph to enable problem-structure aware local search.
MAG-Bot proposes a multi-agent auditing framework for social bot detection that leverages large language models and a graph-structured agent network. It reconstructs graph data into account-level records and compares baselines, finding zero-shot auditing feasible but with recall blind spots, while the multi-agent system improves detection performance.
The paper studies sensitive community inference risk from graph neural networks and considers defenses to conceal or reduce detectable community structures. It analyzes privacy implications in network settings and proposes mechanisms to prevent unwanted disclosure of group-level patterns.