🕸️ Graph Learning¶
🧠 NeurIPS2026 · 44 paper notes
📌 Same area in other venues: 🎞️ ECCV2026 (1) · 📷 CVPR2026 (8) · 🔬 ICLR2026 (118) · 💬 ACL2026 (24) · 🧪 ICML2026 (35) · 🤖 AAAI2026 (37)
🔥 Top topics: GNNs ×6 · Diffusion Models ×6
- Beyond the Golden Teacher: Enhancing Graph Learning Through LLM-GNN Co-Teaching
-
Addressing the flaw where designating either model as a "golden teacher" propagates its blind spots in few-shot graph learning, this paper proposes LLM-GNN Co-Teaching—a bidirectional framework that dynamically exchanges confident pseudo-labels via architecture-specific small-loss criteria and extracts temporal self-correction preference pairs (RPL-PO) across training rounds, setting a new SOTA across six benchmarks.
- Bridging Academia and Industry: A Comprehensive Benchmark for Attributed Graph Clustering
-
Introduces PyAGC, a production-ready attributed graph clustering benchmark and modular library that unifies algorithms under the Encode-Cluster-Optimize framework, breaks the full-batch memory ceiling via mini-batch neighbor sampling, and evaluates 17 methods across 12 diverse datasets spanning up to 111M nodes.
- C_4-Equivariant Flow Matching on Anisotropic Power-Diagram Graphs for Microstructure Generation
-
This paper models polycrystalline microstructures as compact anisotropic power diagrams (APDs) and learns their joint generative distribution using a \(C_4\)-equivariant graph flow matching network, enabling arbitrary pixel-resolution rendering and training-free gradient guidance to synthesize complex metallurgical morphologies such as welds, cast slabs, 3D-printed layers, and heterogeneous lamellar structures.
- Chaining 2-FWL GNNs for Combinatorial Graph Alignment
-
For the purely structural, unlabelled combinatorial graph alignment problem, this paper introduces a sequential chaining procedure of 2-FWL expressive Folklore GNNs with non-differentiable discrete ranking feedback, breaking through 1-WL expressivity limits on regular graphs and dramatically improving FAQ initialization on hard sparse random graphs.
- Chasing Label Shifters: A Change-Aware Framework for Dynamic Graph Node Classification
-
To address the systematic failure of dynamic graph node classification on "label shifters"—nodes whose class labels transition across snapshots and are overwhelmingly misclassified into their previous classes—this paper identifies neighborhood-aggregation-induced "embedding inertia" as the root cause and proposes CHASE, a plug-and-play, model-agnostic change-aware framework that substantially boosts shifter accuracy while preserving performance on stable nodes.
- Concorde: Geometry-Aware Link Prediction via Decoupled Energy Minimization
-
Addressing the spectral instability and symmetry-breaking dilemma when coupling diffusion operators with geometric representations in sparse graphs, Concorde decomposes link likelihood into three decoupled energy branches—a parameter-free spectral anchor, a mixed-curvature residual, and a tangent geodesic steering field—nearly doubling early retrieval precision on sparse networks.
- DiPhon: Diffusion on Graphons for Scalable Graph Generation
-
DiPhon formulates a continuous diffusion process on graphon space via a Jacobi stochastic differential equation (SDE) and derives a moment-aligned discretized process on finite graphs, enabling graph diffusion models trained on small graphs (20–80 nodes) to generate valid larger graphs up to 300 nodes zero-shot without retraining.
- DisRFM: Polar Riemannian Flow Matching for Structure-Preserving Graph Domain Adaptation
-
Addressing structural degeneration and minimax training instability in Euclidean adversarial graph domain adaptation, DisRFM disentangles radial topological scales from angular class semantics on constant-curvature manifolds and achieves structure-preserving transport via topology-conditioned Riemannian flow matching.
- Efficient Dynamic Algorithms for Graph Neural Networks with Non-Linear Propagation
-
The paper maintains nonlinear propagation fixed points on dynamic undirected graphs using residual certificates, endpoint rescaling, and selective pushes, proves deterministic amortized update bounds under degree-normalized error, and shows that nonlinearity can improve static classification accuracy without substantially increasing dynamic propagation cost.
- Efficient Memory Crystallization for Graph Learning under Non-Stationary Distribution Shifts
-
Addressing the prohibitive runtime and memory footprints of generative replay modules in continual test-time graph adaptation under non-stationary drift, EMC distills incoming graphs into compact, low-rank memory graphs in closed form, leveraging a state-evolving field with provably tighter generalization bounds to achieve SOTA accuracy while slashing runtime by 87.4% and GPU memory by 92.4%.
- EMERGE: A Benchmark for Updating Knowledge Graphs with Emerging Textual Knowledge
-
Addressing the limitation of standard information extraction systems that operate independently of graph state, this paper introduces EMERGE, the first benchmark for text-driven knowledge graph updating across 7 yearly snapshots, formalizing five fundamental graph update operations spanning entity minting, link inference, and triple deprecation over 1.45M update operations.
- Equivariant Spherical Transformer for Efficient Molecular Modeling
-
Addressing the limited non-linearity and steep computational costs of Clebsch-Gordan tensor product convolutions in equivariant GNNs, this paper proposes the Equivariant Spherical Transformer (EST), which maps group representations to the continuous spherical spatial domain via uniform Fibonacci lattice sampling, capturing cross-orientation interactions with spherical attention and hybrid experts while provably maintaining strict SO(3) equivariance.
- Flowette: Flow Matching with Graphette Priors for Graph Generation
-
Flowette introduces a continuous flow matching framework for graph generation that integrates graphette structural priors via controlled subgraph edits, topology-aware fused Gromov-Wasserstein coupling, and endpoint consistency regularization to achieve state-of-the-art structural and chemical fidelity.
- Get a Grip, This Will Be a Long Trip: A Quantifiable Long-Range Framework for Verifying Over-squashing
-
This paper introduces four statistical axioms for verifiable long-range graph tasks, constructs the provably sound TRIP/GRIP benchmark suite based on \(f\)-stable distributions on arbitrary topologies, and derives the first closed-form per-range MLE test error lower bound.
- GraDE: A Graph Diffusion Estimator for Frequent Subgraph Discovery in Neural Architectures
-
This paper introduces GraDE, the first framework to repurpose discrete graph diffusion models as subgraph frequency estimators by scoring structural typicality, combining it with beam search to efficiently discover large-scale neural network motifs where traditional sampling collapses.
- Graph Cascades: Contagion-Based Mesoscopic Rewiring for Structure-Aware Graph Machine Learning
-
This paper introduces Graph Cascades, an \(\mathcal{O}(|V|+|E|)\) contagion-based mesoscopic rewiring strategy that promotes node pairs reinforced by multi-path short walks to direct neighbors, overcoming the dichotomy between purely local message passing and dense global attention and delivering massive accuracy gains on heterophilic and moderate- to high-degree homophilic graphs.
- Graph Energy Matching: Transport-Aligned Energy-Based Modeling for Graph Generation
-
This paper introduces Graph Energy Matching (GEM), a discrete generative framework inspired by Jordan–Kinderlehrer–Otto (JKO) transport optimization that learns a permutation-invariant scalar potential to seamlessly unify deterministic greedy transport and Metropolis-Hastings discrete mixing, matching or outperforming state-of-the-art discrete diffusion models on molecular benchmarks.
- Graph Learning Should Move Beyond Restrictive Views of Spectral and Message-Passing GNNs
-
This paper analyzes the fundamental symmetry foundations separating spectral GNNs and MPNNs, proves their expressive and approximation equivalence under linear spectral filtering and local aggregation shift operators, and proposes a generalized eigenbasis-invariant framework with complementary theoretical perspectives.
- Graph Learning Should Move Beyond Restrictive Views of Spectral and Message-Passing Architectures
-
Addressing the long-standing fragmentation between spatial message-passing and spectral graph neural networks, this position paper establishes a rigorous definition of spectral GNNs grounded in eigenbasis invariance, proves that linear spectral filters on aggregation graph shift operators are strictly equivalent to MPNNs in 1-WL separation power and bounded-spectrum approximation, and formalizes spectral positional encodings as an orthogonal design dimension.
- GraphIP–Bench: How Hard Is It to Steal a Graph Neural Network, and Can We Stop It?
-
GraphIP-Bench establishes the first unified black-box benchmark for GNN model extraction and intellectual property defense across 12 attacks, 12 defenses, and 10 graphs, uncovering that training-time parameter-side graph watermarks suffer dramatic signal collapse once the target is extracted into a newly trained surrogate.
- HiGraph: A Large-Scale Hierarchical Graph Dataset for Malware Analysis
-
This paper introduces HiGraph, the first large-scale hierarchical graph dataset for Android malware analysis containing 499K global function call graphs and over 200M nested control flow graphs across a 10-year timeline, demonstrating that intra-procedural cyclomatic complexity is the key discriminative invariant against temporal concept drift.
- HOPSE: Scalable Higher-Order Positional and Structural Encoder for Combinatorial Representations
-
HOPSE eliminates computationally expensive Higher-Order Message Passing (HOMP) layers by decomposing combinatorial complexes into Hasse graphs to extract expressive positional and structural encodings, achieving linear scalability with respect to cell counts while matching or surpassing state-of-the-art predictive performance.
- HoTS: Homophily-Aware Temperature Scaling for Graph Neural Network Calibration
-
HoTS generates positive node-wise temperatures from predictive entropy and local homophily estimated by an auxiliary GCN, using an idealized CSBM inverse-homophily temperature law as a structural prior; it preserves predicted classes and achieves the best mean ECE of 4.79% across 18 datasets, but does not win on every dataset or coverage level.
- Hypergraph Generation via Structured Stochastic Diffusion
-
This paper introduces Hedge, a structured stochastic diffusion model operating directly on relaxed hypergraph incidence matrices by combining a data-adaptive two-sided heat operator with an Ornstein–Uhlenbeck drift, achieving exact conditional reverse-drift targets and superior higher-order structural generation.
- Hypergraph Representation Learning with Hyperlink Random Effects
-
Addressing the challenges of latent formation heterogeneity and missing low-rank structure in general non-uniform hypergraphs, this paper introduces a hypergraph representation learning framework via hyperlink random effects with provable parameter identifiability, statistical consistency guarantees, and closed-form variational EM and score-based diffusion inference algorithms.
- K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs
-
This paper introduces K12-KGraph, the first multimodal curriculum-aligned knowledge graph extracted from official Chinese K-12 textbooks, deriving a 23,640-question structural benchmark K12-Bench and a 7,335-sample SFT corpus K12-Train that systematically expose and resolve the severe lack of curriculum structural understanding in current LLMs.
- MAGNET: Manifold-Aware Graph Diffusion Network for Connectome Generation
-
Addressing mode averaging and clinical heterogeneity loss in functional brain connectome generation caused by coarse clinical labels, MAGNET operates on the normalized-Cholesky valid correlation manifold with an ROI-token graph denoiser, coupling class-anchored Gaussian mixture subtypes, an amortized structural bridge, and relevance-preserving noise modulation to achieve strict geometric validity alongside faithful conditional support recovery.
- Mean-Field Control on Sparse Graphs: From Local Limits to GNNs via Neighborhood Distributions
-
Addressing the failure of classical mean-field control on realistic sparse networks, this paper lifts the system state to probability measures over decorated rooted neighborhoods, rigorously proves horizon-dependent locality, establishes a tractable dynamic programming principle and policy gradient theorem, and theoretically justifies graph neural networks as universal mean-field approximators.
- Network Learning with Semi-Relaxed Gromov-Wasserstein
-
This paper introduces a nonparametric network learning framework using semi-relaxed Gromov-Wasserstein (srGW) barycenters that relaxes NP-hard discrete community assignment to continuous probabilistic couplings, rigorously proving an \(\mathcal{O}(1/n)\) deterministic optimality gap and achieving minimax-optimal convergence rates for both SBMs and Hölder graphons.
- Neural Proposals, Symbolic Guarantees: Neuro-Symbolic Graph Generative Modeling
-
Introduces NSGGM, a neuro-symbolic graph generative framework that decouples generation into autoregressive neural motif proposals and SMT-based symbolic assembly, delivering verifiable correctness-by-construction under strict compositional logical rules while retaining high distributional fidelity.
- Neural Structural Reasoner: A Brain-inspired Architecture for Reasoning over Structured Knowledge
-
NSR binds entities, relations, and relation chains to readable network units and connections, performing link prediction through relation-association retrieval and explicit graph traversal; it achieves an MRR of 0.8142 on Nations, but its advantages are dataset-dependent, and the efficiency of its accelerated static implementation should not be equated with that of its online brain-inspired dynamics.
- OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data
-
This paper introduces OgBench, the first end-to-end modular benchmarking framework for graph-level prediction in the low-sample, high-node (\(n \ll p\)) omics regime; an extensive evaluation across 80,000 trained models reveals that mainstream GNNs rarely outperform simple MLPs and linear baselines, prompting a critical reassessment of biological graph inductive biases.
- Orlicz–Sobolev with Musielak: An Efficient Regularization Approach for Graph-Based Semi-Supervised Learning
-
To overcome the rigid \(L^p\) geometric coupling and severe computational bottleneck of graph-based Sobolev Integral Probability Metrics (Sobolev IPM), this paper introduces generalized Sobolev IPM (GSI) via Orlicz–Sobolev spaces and proves an equivalence to Musielak gradient norms, reducing the functional optimization problem to a highly efficient univariate convex program.
- Oversmoothing as Representation Degeneracy in Neural Sheaf Diffusion
-
This paper develops a quiver-theoretic framework identifying cellular sheaves with incidence-quiver representations, interprets oversmoothing as representation degeneracy, and introduces moment-map regularizers alongside symmetry-breaking rectangular stalks to prevent trivial harmonic collapse.
- PANDA: Prior-guided Attentional Dual-path Architecture
-
Addressing the issues where fixed structural priors in hypergraph neural networks propagate misleading heterophilic signals and receiver-normalized attention excessively dilutes multi-sender contributions through zero-sum competition, this paper proposes PANDA, a modular plug-in combining prior-guided main-path interpolation with a non-competitive gated auxiliary path that significantly boosts node classification across directed and undirected settings.
- Pointwise Lipschitz Continuous Graph Algorithms
-
This paper introduces a general Proximal Gradient Trajectory Analysis (PGTA) framework that unlocks the use of weighted regularizers in convex relaxations, establishing an asymptotically tight pointwise Lipschitz continuous algorithm for minimum vertex SS-TT cut and extending stable algorithms to general bipartite \(b\)-matching and packing integer programs.
- Rank Is Not Capacity: Spectral Occupancy for Latent Graph Models
-
Addressing the flaw where latent graph models treat latent dimension as a rigid, brittle hyperparameter detached from effective representational capacity, Spectra measures and controls capacity end-to-end via the Shannon effective rank of a trace-normalized positive semidefinite kernel, deterministically producing optimal nested spectral prefixes from a single training run.
- Relation-Aware Graph Foundation Model
-
REEF treats transferable relation semantics as the unit of graph foundation modeling and uses semantic hypernetworks to generate aggregators, classifiers, and dataset projectors, reaching 48.66%, 48.65%, and 47.72% accuracy on three target graphs under 1-shot classification, while retaining clear limitations for unseen relations, zero-shot tasks, and computational cost.
- Spectral Graph Sparsification Preserves Representation Geometry in Graph Neural Networks
-
This paper establishes a representation-geometric theory of spectral graph sparsification for graph neural networks (GNNs), proving that any \(\varepsilon\)-spectral sparsifier induces at most \(O(\varepsilon)\) perturbations across polynomial graph filters, multilayer hidden embeddings, Gram matrices, and finite-time gradient descent parameter trajectories, while empirically validating the preservation of pairwise distances, local neighborhoods, and class-centroid geometry via effective-resistance sparsification.
- Spectral Reversal: Counteracting Singular Value Bias for Graph Prompting
-
SRP preferentially reads weak directions through singular-value soft masks at each frozen GNN layer, supplements features discarded by the original weights through null-space PCA, and adds low-rank prompts after aggregation; it achieves 68.11% accuracy on 5-shot Cora with GraphCL pretraining, but does not outperform the strongest baseline on every dataset.
- Toward a Unified Statistical Theory of Unsupervised Pretraining and Supervised Neural Knowledge Graph Learning
-
This paper proposes the Pretrained Neural Knowledge Graph (PNKG) two-stage framework, which fuses heterogeneous unlabeled entity features via weighted multi-view kernel principal component analysis (KPCA) and predicts triples using relation-specific bounded ReLU networks, establishing the first nonasymptotic end-to-end statistical theory that disentangles unsupervised pretraining subspace recovery error from downstream supervised risk and achieving consistent gains on WordNet and PrimeKG.
- Understanding and Mitigating Under-Confidence in GNNs from the Final Layer
-
SCAR explains GNN under-confidence through the final classifier, reduces only final-layer weight decay during training, and moves representations toward predicted-class prototypes at inference time using training-set adjacency groups, reducing ECE across multiple node classification settings without guaranteeing unchanged labels or accuracy.
- Weisfeiler and Leman Follow the Arrow of Time: Expressive Power of Message Passing in Temporal Event Graphs
-
Addressing the limitation of existing temporal GNNs that neglect the causal topology shaped by the directed arrow of time, this paper introduces time-respecting path isomorphism, establishes its equivalence to static graph isomorphism on augmented event graphs, and proves that directed message passing matches the expressive power of the directed 1-WL test.
- When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors
-
Addressing the fixed-point pathology in featureless discrete graph diffusion caused by edge independence and exchangeability enforcing an Erdős–Rényi prior, this paper proposes the Latent Sociability Prior (LSP) and a joint diffusion framework that co-evolves continuous node sociabilities with discrete edges without complex encoders, significantly boosting higher-order topology generation.