🕸️ Graph Learning¶

🧠 NeurIPS2025 · 52 paper notes

Agint: Agentic Graph Compilation for Software Engineering Agents: This paper proposes Agint, a graph compiler that progressively compiles natural language intent into typed DAGs through a six-level type floor (TEXT→TYPED→SPEC→STUB→SHIM→PURE), paired with a hybrid JIT runtime and a Unix-style toolchain. This transforms AI code generation from brittle single-pass text prediction into a structured, parallelizable, and reproducible compilation process.
BLISS: Bandit Layer Importance Sampling Strategy for Efficient Training of Graph Neural Networks: This paper proposes BLISS, which formulates layer-wise neighbor sampling in GNNs as a multi-armed bandit problem. Using the EXP3 algorithm, it dynamically adjusts per-edge sampling probabilities with the variance contribution of neighbors to node representations as the reward signal, achieving accuracy on par with or exceeding full-batch training on GCN and GAT.
Deliberation on Priors: Trustworthy Reasoning of Large Language Models on Knowledge Graphs: This paper proposes DP (Deliberation on Priors), a framework that leverages structural priors from knowledge graphs via progressive knowledge distillation to generate faithful relational paths, and validates reasoning reliability through a reasoning introspection strategy based on constraint priors, achieving new state-of-the-art performance on KGQA benchmarks.
Diagnosing and Addressing Pitfalls in KG-RAG Datasets: Toward More Reliable Benchmarking: A systematic audit of 16 KGQA datasets reveals an average factual correctness of only 57% (WebQSP: 52%, MetaQA: 20%). The paper proposes KGQAGen, a framework that constructs high-quality multi-hop QA datasets via LLM-guided subgraph expansion and automatic SPARQL validation, yielding KGQAGen-10k with 96.3% accuracy. The study further demonstrates that the primary bottleneck in KG-RAG lies in retrieval rather than reasoning.
DuetGraph: Coarse-to-Fine Knowledge Graph Reasoning with Dual-Pathway Global-Local Fusion: DuetGraph proposes a dual-pathway (message passing + global attention) parallel fusion model and a coarse-to-fine reasoning optimization strategy. By separating rather than stacking local/global information processing, it mitigates score over-smoothing in KG reasoning, achieving SOTA on both inductive and transductive tasks with up to 8.7% MRR improvement and 1.8× training speedup.
Dynamic Bundling with Large Language Models for Zero-Shot Inference on Text-Attributed Graphs: DENSE proposes a "text bundling" strategy that packages textually and topologically/semantically similar nodes into bundles, queries LLMs for bundle-level labels, supervises GNN training via entropy-based and ranking-based losses, and dynamically refines bundles to exclude noisy nodes. It achieves comprehensive zero-shot inference improvements over GPT-4o and graph foundation models across 10 TAG datasets.
Elastic Weight Consolidation for Knowledge Graph Continual Learning: An Empirical Evaluation: This paper systematically evaluates Elastic Weight Consolidation (EWC) for continual learning of TransE knowledge graph embeddings on FB15k-237, finding that EWC reduces catastrophic forgetting from 12.62% to 6.85% (a 45.7% reduction), and reveals that task partitioning strategy (relation-based vs. random) has a substantial impact on forgetting metrics (a difference of 9.8 percentage points).
ESCA: Contextualizing Embodied Agents via Scene-Graph Generation: This paper proposes the ESCA framework, which provides structured visual understanding context for MLLM-driven embodied agents via open-vocabulary scene graph generation (the SGClip model), substantially reducing perception error rates and improving task completion rates.
FALCON: An ML Framework for Fully Automated Layout-Constrained Analog Circuit Design: FALCON proposes an end-to-end framework for automated analog/RF circuit design via a three-stage pipeline: MLP-based topology selection, edge-centric GNN performance prediction, and differentiable layout-constrained gradient inference. Trained on a million-scale Cadence simulation dataset, the framework achieves >99% topology selection accuracy, <10% performance prediction error, and sub-second per-instance inference.
FastJAM: a Fast Joint Alignment Model for Images: FastJAM is a graph-based fast joint image alignment method that computes pairwise keypoint correspondences using off-the-shelf image matchers, constructs a keypoint graph via fast non-parametric clustering, employs a GNN to propagate and aggregate information for predicting per-image homography parameters, and adopts an inverse-compositional loss to eliminate the need for regularization hyperparameters. It reduces joint alignment time from hours/minutes to approximately 49 seconds while achieving alignment quality superior to or on par with existing methods.
From Sequence to Structure: Uncovering Substructure Reasoning in Transformers: This paper presents empirical and theoretical analyses revealing how decoder-only Transformers understand graph structure from text sequences. It proposes "Induced Substructure Filtration" (ISF) to explain the layer-wise substructure identification mechanism, and extends this framework to validate consistency in LLMs, support compositional graph reasoning (Thinking-in-Substructures), and enable substructure extraction in attributed graphs (molecular graphs).
Generative Graph Pattern Machine: This paper proposes the Generative Graph Pattern Machine (G2PM), a fully message-passing-free generative Transformer framework for graph pre-training. It tokenizes graph instances (nodes/edges/graphs) into substructure sequences via random walks and performs self-supervised pre-training under a Masked Substructure Modeling objective. G2PM comprehensively outperforms existing graph pre-training methods on node/link/graph classification and cross-domain transfer tasks, while exhibiting model and data scaling laws analogous to those observed in NLP and CV.
Geometric Imbalance in Semi-Supervised Node Classification: This work formally introduces the concept of "geometric imbalance" in semi-supervised node classification for the first time—showing that message passing on class-imbalanced graphs causes minority-class nodes to exhibit geometric ambiguity in Riemannian manifold embedding spaces—and proposes the UNREAL framework to systematically address this issue via three modules: dual-path pseudo-label alignment, node reordering, and geometric imbalance sample discarding.
GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation: This paper proposes GFM-RAG, the first graph foundation model-driven retrieval-augmented generation framework, which performs single-pass multi-hop reasoning over knowledge graphs via a query-dependent GNN. With only 8M parameters, GFM-RAG achieves zero-shot generalization to unseen datasets and substantially outperforms state-of-the-art methods on multi-hop QA retrieval benchmarks.
GnnXemplar: Exemplars to Explanations -- Natural Language Rules for Global GNN Interpretability: This paper proposes GnnXemplar, a framework grounded in the cognitive-science Exemplar Theory. It selects representative nodes (exemplars) in the GNN embedding space and employs an LLM with iterative self-refinement to generate natural-language Boolean rules, achieving global interpretability for node-classification GNNs on large-scale graphs.
Graph Neural Networks for Efficient AC Power Flow Prediction in Power Grids: This work models power networks as graph structures (buses as nodes, transmission lines as edges) and investigates four GNN architectures — GCN, GAT, SAGEConv, and GraphConv — for predicting AC power flow solutions (voltage magnitudes and phase angles). Experiments on IEEE 14/30/57/118-bus test systems demonstrate that GNNs can efficiently substitute traditional Newton-Raphson solvers.
Graph Neural Networks for Interferometer Simulations: This work presents the first application of graph neural networks to optical interferometer simulation, employing a GATv2 + KAN architecture to predict electromagnetic field power and spatial intensity distributions within LIGO interferometers. The approach achieves inference speeds up to 815× faster than the standard simulation software (FINESSE) while maintaining satisfactory physical accuracy.
Graph Persistence goes Spectral: This paper proposes SpectRe — a novel topological descriptor that incorporates graph Laplacian spectral information into persistent homology (PH) graphs. It proves that SpectRe is strictly more expressive than either PH or spectral methods alone, establishes a local stability theory, and demonstrates improved GNN graph classification performance on both synthetic and real-world datasets.
GraphFaaS: Serverless GNN Inference for Burst-Resilient, Real-Time Intrusion Detection: This paper proposes GraphFaaS, a serverless inference architecture for GNN-based intrusion detection. Through incremental provenance graph construction, feature-length-aware parallel node embedding, and greedy best-fit subgraph partitioning, GraphFaaS reduces mean detection latency from 14.16 seconds to 2.1 seconds (6.7×) and the coefficient of variation from 1.46 to 0.52 (64% reduction), maintaining stable low latency under bursty workloads without sacrificing detection accuracy.
GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks: This paper proposes GraphTOP, the first graph topology-oriented prompting framework, which formulates topology-oriented prompting as an edge rewiring problem and relaxes it into a continuous space via Gumbel-Softmax. GraphTOP outperforms six baseline methods across five datasets and four pre-training strategies.
Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems: This paper proposes the Heterogeneous Swarms algorithm, which models multi-LLM systems as directed acyclic graphs (DAGs) and employs particle swarm optimization (PSO) to jointly optimize model roles (graph topology) and model weights, achieving an average improvement of 18.5% over 17 baselines across 12 tasks.
Interaction-Centric Knowledge Infusion and Transfer for Open-Vocabulary Scene Graph Generation: This paper proposes ACC, an interaction-centric framework that addresses the critical matching problem in open-vocabulary scene graph generation (OVSGG) by shifting from the conventional object-centric paradigm to an interaction-driven one. During the knowledge infusion stage, bidirectional interaction prompts are used to generate more accurate pseudo supervision; during the knowledge transfer stage, interaction-guided query selection and interaction-consistency knowledge distillation reduce mismatches. ACC achieves state-of-the-art performance on three benchmarks: VG, GQA, and PSG.
Learning Repetition-Invariant Representations for Polymer Informatics: This paper proposes GRIN (Graph Repetition-Invariant Network), which achieves invariance to the number of repeated monomer units in polymer representations via Max aggregation and a specialized graph construction strategy, addressing a fundamental symmetry problem in polymer representation learning.
Logical Expressiveness of Graph Neural Networks with Hierarchical Node Individualization: This paper proposes Hierarchical Ego GNNs (HEGNNs), which generalize subgraph GNNs through a hierarchical node individualization mechanism, forming a hierarchy of models with strictly increasing expressive power. On bounded-degree graphs, the paper proves that the distinguishing power of HEGNN node classifiers is equivalent to graded hybrid logic (\(\mathcal{HL}_k\)), thereby unifying the expressiveness analysis of various GNN variants.
Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness-Generalization Perspective: This paper theoretically reveals the smoothness-generalization dilemma inherent in GNN message passing, and proposes the IGNN framework with three minimal design principles — separative neighborhood transformation, inceptive aggregation, and neighborhood relationship learning — to systematically alleviate this dilemma. IGNN achieves top performance among 30 baselines and demonstrates universality across both homophilic and heterophilic graphs.
Moscat: Mixture of Scope Experts at Test for Generalizing Deeper GNNs: Grounded in PAC-Bayes generalization theory, this paper proves that varying GNN depth induces generalization preference drift across node subgroups with different homophily levels. It proposes Moscat—a post-processing attention-gating model that adaptively fuses independently trained GNN experts of different depths at test time on a per-node basis—achieving significant improvements across diverse GNN architectures and datasets.
MoEMeta: Mixture-of-Experts Meta Learning for Few-Shot Relational Learning: This paper proposes MoEMeta, a framework that employs a Mixture-of-Experts model to learn globally shared relational prototypes for cross-task generalization, combined with a task-customized projection adaptation mechanism to capture local context, achieving state-of-the-art performance on three KG benchmarks.
Nonlinear Laplacians: Tunable Principal Component Analysis under Directional Prior Information: This paper proposes a nonlinear Laplacian spectral algorithm that fuses spectral information with directional prior information by adding a diagonal matrix—obtained by applying a nonlinear function \(\sigma\) to the degree vector of the observation matrix \(\bm{Y}\)—to \(\hat{\bm{Y}}\). The approach significantly reduces the signal detection threshold in the biased sparse PCA problem (from \(\beta^*=1\) to approximately \(0.76\)).
OCN: Effectively Utilizing Higher-Order Common Neighbors for Better Link Prediction: This paper identifies redundancy and over-smoothing issues in higher-order common neighbors (CN) for link prediction, and proposes orthogonalization (Gram-Schmidt to remove inter-order linear dependence) combined with normalization (dividing by path count, a generalized resource allocation heuristic) as a solution. The method achieves an average improvement of 7.7% in HR@100 across 7 datasets, with a 13.3% gain on the DDI dataset.
Over-squashing in Spatiotemporal Graph Neural Networks: This paper provides the first formal treatment of over-squashing in spatiotemporal graph neural networks (STGNNs), uncovering a counterintuitive "temporal sink" phenomenon in causal convolutions—whereby the earliest timestep exerts the greatest influence on the final representation—and proves that time-and-space (T&S) and time-then-space (TTS) architectures are equivalent in terms of information bottlenecks, offering theoretical justification for the computationally efficient TTS design.
P-DRUM: Post-hoc Descriptor-based Residual Uncertainty Modeling for Machine Learning Potentials: This paper proposes P-DRUM, a simple and efficient post-hoc uncertainty quantification framework that leverages descriptors from a trained graph neural network potential to estimate prediction residuals as uncertainty proxies, requiring no modification to the original model architecture or training pipeline.
Practical Bayes-Optimal Membership Inference Attacks: This paper proposes BASE and G-BASE, two practical Bayes-optimal membership inference attack methods targeting i.i.d. data and graph-structured data, respectively, achieving theoretical optimality while substantially reducing computational cost.
PKD: Preference-driven Knowledge Distillation for Few-shot Node Classification: PKD is a framework that jointly leverages LLMs and multiple GNN teachers for few-shot node classification on text-attributed graphs. A GNN-preference node selector (GNS) uses KL divergence-based uncertainty to identify nodes requiring LLM annotation, while a node-preference GNN selector (NGS) employs RL to match each node with its optimal GNN teacher. PKD achieves consistent state-of-the-art performance across 9 datasets (e.g., Cornell 87% vs. baselines 59–82%).
Principled Data Augmentation for Learning to Solve Quadratic Programming Problems: This paper proposes a principled data augmentation framework based on affine transformations of the KKT system, generating optimality-preserving augmented instances for MPNN-based learning-to-optimize (L2O) on linear programming (LP) and quadratic programming (QP) tasks. Combined with contrastive pretraining, the framework substantially improves performance in data-scarce and out-of-distribution (OOD) generalization settings.
Reasoning Meets Representation: Envisioning Neuro-Symbolic Wireless Foundation Models: This paper proposes a vision framework that integrates the neuro-symbolic (NeSy) paradigm with Wireless Physical-layer Foundation Models (WPFMs)—employing WPFMs as a neural perception engine to generate RF embedding vectors, and an ontology-driven knowledge graph together with a differentiable logic layer as the symbolic reasoning component. The resulting system achieves interpretable, generalizable, and compliance-verifiable wireless AI, providing a concrete technical pathway toward AI-native 6G networks.
Relieving the Over-Aggregating Effect in Graph Transformers: This paper identifies the over-aggregating phenomenon in Graph Transformers—wherein a large number of nodes are aggregated with near-uniform attention scores, diluting critical information—and proposes Wideformer, which alleviates this issue via divided aggregation and guided attention. As a plug-and-play module, Wideformer consistently improves backbone model performance across 13 datasets.
ReMindRAG: Low-Cost LLM-Guided Knowledge Graph Traversal for Efficient RAG: This paper proposes ReMindRAG, a KG-RAG system that combines LLM-guided KG traversal (node exploration + exploitation) with a training-free memory replay mechanism. The system stores LLM traversal experience in edge embeddings, enabling significant reduction in LLM calls for subsequent similar queries (~50% cost reduction) while improving answer accuracy (5%–10% gain).
Self-Supervised Discovery of Neural Circuits in Spatially Patterned Neural Responses with Graph Neural Networks: A GNN-based self-supervised framework is proposed that infers latent synaptic connectivity via a structure learning module while simultaneously predicting future spiking activity via a spike prediction module. The approach substantially outperforms statistical inference baselines on both simulated ring attractor network data and real mouse head-direction cell recordings.
Sketch-Augmented Features Improve Learning Long-Range Dependencies in Graph Neural Networks: This paper proposes Sketched Random Features (SRF), which injects random kernel-space projections of node features into every layer of a standard message-passing GNN, simultaneously alleviating oversquashing, oversmoothing, and limited expressiveness, with rigorous theoretical guarantees and low computational overhead.
S'MoRE: Structural Mixture of Residual Experts for Parameter-Efficient LLM Fine-tuning: This paper proposes S'MoRE, a framework that organizes low-rank residual experts into a multi-layer tree structure and constructs token-specific "residual trees" via hierarchical routing, achieving exponentially growing structural flexibility with parameter counts comparable to LoRA, thereby substantially improving LLM fine-tuning performance.
Solar-GECO: Perovskite Solar Cell Property Prediction with Geometric-Aware Co-Attention: This paper proposes Solar-GECO, a multimodal framework that encodes the 3D crystal structure of the perovskite absorber layer via a geometric GNN and the remaining device layers via LLM text embeddings, fuses them through a co-attention module, and predicts power conversion efficiency (PCE) along with its uncertainty, reducing MAE from 3.066 to 2.936.
Spatio-Temporal Directed Graph Learning for Account Takeover Fraud Detection: This paper proposes ATLAS, a framework that reformulates account takeover (ATO) fraud detection as a node classification problem on spatio-temporal directed graphs. By constructing causal directed graphs via temporal windows and nearest-neighbor constraints, and combining lag-aware label propagation with a GraphSAGE encoder, ATLAS achieves a +6.38% AUC improvement and over 50% reduction in user friction on a production graph at Capital One with 100M nodes and 1B edges.
SPOT-Trip: Dual-Preference Driven Out-of-Town Trip Recommendation: This paper proposes SPOT-Trip, the first framework to systematically address out-of-town trip recommendation. By integrating knowledge graph-enhanced static preference learning, neural ODE-driven dynamic preference learning, and a preference fusion module, the framework achieves up to 17.01% improvement over state-of-the-art baselines on two real-world datasets.
SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs: This paper proposes SSTAG, which jointly distills complementary knowledge from LLMs and GNNs into a structure-aware MLP via dual knowledge distillation, and incorporates a memory bank mechanism to store prototype representations, enabling efficient and scalable cross-domain self-supervised pre-training on text-attributed graphs.
TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction: This paper is the first to systematically identify the heterogeneity problem in temporal graph interactions (interaction intervals follow a power-law distribution), and proposes the TAMI framework comprising two modules—Log Time Encoding (LTE) and Link History Aggregation (LHA)—that can be seamlessly integrated into existing TGNNs, consistently improving link prediction performance across 16 datasets with gains of up to 87.05%.
The Underappreciated Power of Vision Models for Graph Structural Understanding: This paper reveals the severely underappreciated capability of vision models (ResNet/ViT/Swin, etc.) for graph structural understanding. By rendering graphs as images and processing them with visual encoders, these models significantly outperform GNNs in global topology perception and cross-scale generalization. The paper also introduces the GraphAbstract benchmark to systematically evaluate this finding.
Uncertain Knowledge Graph Completion via Semi-Supervised Confidence Distribution Learning: ssCDL converts triple confidence scores from scalars into Gaussian confidence distributions to capture supervisory signals from neighboring confidence values, and employs meta self-training to generate high-quality pseudo confidence labels for negatively sampled triples, thereby rebalancing the training data. The method significantly outperforms all baselines on both confidence prediction and link prediction for uncertain knowledge graph completion.
Unifying and Enhancing Graph Transformers via a Hierarchical Mask Framework: This paper proposes a unified hierarchical mask framework that reveals the equivalence between Graph Transformer architectures and attention masks, and introduces M3Dphormer, which achieves efficient adaptive modeling of local/cluster/global interactions via multi-level masks, bi-level expert routing, and a dual attention computation scheme, achieving state-of-the-art results on 9 benchmarks.
Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with LLMs: This paper proposes the Cross framework, which employs LLMs to dynamically summarize the semantic evolution of node neighborhoods at strategically sampled temporal points (Temporal Reasoning Chain), then bidirectionally fuses text semantics and graph structural temporal information via a semantic-structural co-encoder. The approach achieves an average MRR improvement of 24.7% on temporal link prediction and a 3.7% AUC gain on an industrial dataset (WeChat).
Wavy Transformer: This paper establishes a formal equivalence between Transformer attention layers and graph neural diffusion on complete graphs, and proposes the Wavy Transformer based on second-order wave equations. By exploiting energy conservation properties, the method mitigates over-smoothing in deep Transformers and achieves consistent improvements across NLP, CV, and sparse graph tasks.
What Expressivity Theory Misses: Message Passing Complexity for GNNs: This paper critiques the binary expressivity theory of GNNs for its inability to explain practical performance differences, and proposes MPC—a continuous, task-specific complexity measure grounded in probabilistic lossyWL—achieving a Spearman correlation of -1 with accuracy (versus ρ = 0 for classical WLC), and successfully explaining why GCN with virtual nodes outperforms higher-expressivity higher-order models on long-range tasks.
When No Paths Lead to Rome: Benchmarking Systematic Neural Relational Reasoning: This paper introduces the NoRA benchmark, which systematically breaks the assumption underlying existing relational reasoning benchmarks that "reasoning can be reduced to path composition." By introducing off-path reasoning, ambiguous facts, and multi-relational settings, it reveals fundamental deficiencies in all existing models—including o3—on off-path reasoning.