已确认:2025.md 与 publications_iclr25/README.md 为同一份 ICLR 2025 论文清单(255 篇),子目录 README 提供 BibTeX 与 OpenReview 链接。现有材料充足,可成文。
已确认:2025.md 与 publications_iclr25/README.md 为同一份 ICLR 2025 论文清单(255 篇),子目录 README 提供 BibTeX 与 OpenReview 链接。现有材料充足,可成文。
2025 年 ICLR 图深度学习论文全景解读:Scalability、表达力与语言模型融合的 255 篇前沿工作
2025 年的 ICLR(第十三届 ICLR 会议)收录了大量图深度学习(Graph-based Deep Learning)方向的论文,本文以当前开源仓库 conference-publications/folders/submissions_iclr/2025.md 为骨架,对照 conference-publications/folders/years/2025/publications_iclr25/README.md 的完整分类清单(共 255 篇)与 255 个独立论文子目录(每个目录含 BibTeX 引用格式与 OpenReview 链接),系统梳理 ICLR 2025 上围绕 GNN 的 Scalability、异质性(Heterophily)、过平滑/过挤压、长程依赖、谱方法、随机游走、图 Transformer、超图、拓扑深度学习、表达力、鲁棒性、图语言模型、分子与蛋白质、等变网络、动态/时序图、组合优化等核心议题的代表性工作。读完后,你将获得一张可直接用于论文检索、引用与选题追踪的 ICLR 2025 图学习研究路线图。
这份清单在仓库中如何组织
本仓库(graph-based-deep-learning-literature)以“按会议、按年份、按主题分类”的方式组织图深度学习文献链接。与本文直接相关的资源有以下三处:
- submissions_iclr/2025.md:ICLR 2025 提交/录用论文的主题速览清单,按 30+ 个研究主题分组;
- years/2025/publications_iclr25/README.md:同一批论文的更完整收录版本(255 篇),主题分类更细,并额外包含 Curvature、Knowledge Distillation、Graph Diffusion、Graph Matching、Time Series、Circuits、Proteins、RNA 等补充分类;
- years/2025/publications_iclr25/ 下 255 个子目录:每个目录对应一篇论文,内含可直接复制的 BibTeX 引用条目与指向 OpenReview 的链接,例如 stgcnd_iclr25/README.md、musegnn_iclr25/README.md、uvn_iclr25/README.md。
需要说明:submissions_iclr/2025.md 与 publications_iclr25/README.md 收录的是同一届 ICLR 2025 的图学习论文,但后者分类更全、条目更多(前者约 170 条,后者 255 条),且每个条目都带仓库内链,因此以下解读以两者互补为准。
Scalability:让 GNN 在大图上"可收敛、可证明"
ICLR 2025 的图学习论文中,Scalability(可扩展性)是最受关注的主题之一,共 8 篇工作(完整版清单)或 6 篇(速览版清单),覆盖了从层设计、图稀疏化到数据集压缩的多个层次:
- MuseGNN: Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy:通过最小化基于采样的能量函数来构造可扩展且可收敛的 GNN 层,对应仓库条目 musegnn_iclr25;
- Scalable and Certifiable Graph Unlearning: Overcoming the Approximation Error Barrier:面向图数据删除请求的"可扩展且可认证"的图遗忘学习,对应 scalegun_iclr25;
- ST-GCond: Self-supervised and Transferable Graph Dataset Condensation:自监督、可迁移的图数据集压缩,对应 stgcnd_iclr25;
- Graph Sparsification via Mixture of Graphs:用"图的混合"做图稀疏化,对应 mog_iclr25;
- Accurate and Scalable Graph Neural Networks via Message Invariance:通过消息不变性同时保证准确性与可扩展性,对应 top_iclr25;
- Node Identifiers: Compact, Discrete Representations for Efficient Graph Learning:用紧凑的离散节点标识符提升图学习效率,对应 nodeid_iclr25;
- 完整清单还补充了 Bonsai: Gradient-free Graph Condensation for Node Classification(bonsai_iclr25)与 RTop-K: Ultra-Fast Row-Wise Top-K Selection for Neural Network Acceleration on GPUs(rtopk_iclr25)。
注:速览版与完整版对同一论文的链接会指向同一 OpenReview 论坛。上文中未在 2025.md 中出现、仅在完整版 README 收录的论文,链接均指向完整版对应子目录。
从理论到实践:表达力(Expressivity)研究的深化
表达力是 ICLR 图学习论文的"常青树",2025 年共有 8 篇相关论文,并且进一步细分为"同态表达力"(Homomorphism Expressivity)与"表达力 × 泛化"两个子主题:
Expressivity(8 篇)
- Is uniform expressivity too restrictive? Towards efficient expressivity of GNNs(uetr_iclr25)
- Rethinking the Expressiveness of GNNs: A Computational Model Perspective
- Diss-l-ECT: Dissecting Graph Data with local Euler Characteristic Transforms
- On the Hölder Stability of Multiset and Graph Neural Networks(hsmsgnn_iclr25)
- Towards Stable, Globally Expressive Graph Representations with Laplacian Eigenvectors
- Towards a Complete Logical Framework for GNN Expressiveness(lfgnnexp_iclr25)
- Expressive Power of Graph Neural Networks for (Mixed-Integer) Quadratic Programs
- Graph Neural Networks Can (Often) Count Substructures(gnncss_iclr25)
Homomorphism Expressivity(同态表达力)
- Homomorphism Counts as Structural Encodings for Graph Learning(mose_iclr25)
- Homomorphism Expressivity of Spectral Invariant Graph Neural Networks(hesignn_iclr25)
Expressivity, Generalisation(表达力与泛化的桥接)
- Generalization, Expressivity, and Universality of Graph Neural Networks on Attributed Graphs(geu_iclr25)
- Towards Bridging Generalization and Expressivity of Graph Neural Networks(genexpgnn_iclr25)
这批工作把经典 WL 测试、同态计数、谱不变量、计算模型等理论工具与泛化界结合,表明 ICLR 2025 对"GNN 到底能区分多少图结构"这一问题的研究已经进入精细化阶段——不仅问"能否表达",还问"能否高效表达、稳定表达、泛化表达"。
结构学习的两大顽疾:Oversmoothing 与 Oversquashing
Oversmoothing(过平滑):堆叠多层后节点表示趋于一致。2025 年有 2 篇(速览版)/3 篇(完整版)工作,其中 Residual Connections and Normalization Can Provably Prevent Oversmoothing in GNNs(graphnormv2_iclr25)从理论层面证明残差连接与归一化可以防止过平滑。
Oversquashing(过挤压):长距离信息在瓶颈结构中丢失。代表性工作包括 Understanding Virtual Nodes: Oversquashing and Node Heterogeneity(uvn_iclr25)与 Demystifying Topological Message-Passing with Relational Structures: A Case Study on Oversquashing in Simplicial Message-Passing(tmprs_iclr25)。
Long-range Dependencies(长程依赖) 与之直接相关:
- Port-Hamiltonian Architectural Bias for Long-Range Propagation in Deep Graph Networks(phdgn_iclr25)
- GLoRa: A Benchmark to Evaluate the Ability to Learn Long-Range Dependencies in Graphs(glora_iclr25)
- 完整版另收录 Greener GRASS: Enhancing GNNs with Encoding, Rewiring, and Attention(grass_iclr25)
Heterophily(异质性):当"邻居不同类"时 GNN 失效,2025 年相关论文有 Decoupled Graph Energy-based Model for Node Out-of-Distribution Detection on Heterophilic Graphs(degem_iclr25)与 Let Your Features Tell The Differences: Understanding Graph Convolution By Feature Splitting(tfigfs_iclr25)。
谱方法、随机游走与图 Transformer
Spectral Methods(谱方法):
- Joint Graph Rewiring and Feature Denoising via Spectral Resonance(jdr_iclr25)
- Spectral Graph Coarsening Using Inner Product Preservation and the Grassmann Manifold
- GRASP: Generating Graphs via Spectral Diffusion
- When Graph Neural Networks Meet Dynamic Mode Decomposition(dmdgnn_iclr25)
Random Walks(随机游走):
- Revisiting Random Walks for Learning on Graphs(rwnn_iclr25)
- Learning Long Range Dependencies on Graphs via Random Walks(neuralwalker_iclr25)
Graph Transformers(图 Transformer):
- Graph Transformers Dream of Electric Flow
- Unleashing Graph Transformers with Green and Martin Kernels
超图、拓扑深度学习与高阶结构
Hypergraphs(超图):
- Training-Free Message Passing for Learning on Hypergraphs(tfhnn_iclr25)
- Link Prediction with Relational Hypergraphs
- 完整版另收录 DistillHGNN(distillhgnn_iclr25)、HyperPLR(hyperplr_iclr25)等
Topological Deep Learning(拓扑深度学习):
- TopoTune: A Framework for Generalized Combinatorial Complex Neural Networks
- Topological Blindspots: Understanding and Extending Topological Deep Learning Through the Lens of Expressivity(mcn_iclr25)
- E(n) Equivariant Topological Neural Networks(etnn_iclr25)
- 完整版另收录 MANTRA: The Manifold Triangulations Assemblage(mantra_iclr25)与 Topological Schrodinger Bridge Matching(tsbm_iclr25)
安全性:鲁棒性、投毒、OOD 与异常检测
ICLR 2025 对图模型的可靠性投入了大量篇幅,形成了 Robustness、Poisoning、Out-of-Distribution、Anomaly Detection 四条线索:
Robustness(鲁棒性)
- Towards Generalization Bounds of GCNs for Adversarially Robust Node Classification
- Robustness Inspired Graph Backdoor Defense
- Learning Randomized Algorithms with Transformers
- Certified Defense on the Fairness of Graph Neural Networks
Poisoning(数据投毒)
- Exact Certification of (Graph) Neural Networks Against Label Poisoning
- Provable Robustness of (Graph) Neural Networks Against Data Poisoning and Backdoors
Out-of-Distribution(分布外检测)
- GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation
- STOP! A Out-of-Distribution Processor with Robust Spatiotemporal Interaction
- Spreading Out-of-Distribution Detection on Graphs
Anomaly Detection(异常检测)
- Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark
- SpaceGNN: Multi-Space Graph Neural Network for Node Anomaly Detection with Extremely Limited Labels
这些工作把"对抗攻击下 GCN 的泛化界""标签投毒的精确认证""后门防御""公平性认证"等问题统一纳入图学习的可信框架,与完整版清单中的 Adversarial Attacks 主题相互呼应。
语言模型 × 图:LLM 与图学习的交叉地带
2025 年语言模型与图的交叉是绝对热点,完整版清单中出现了"Graph Language Models"与"Language Models"两个分类,共 9 篇以上工作:
- GOFA: A Generative One-For-All Model for Joint Graph Language Modeling:将图与语言统一进一个生成式模型;
- Scale-Free Graph-Language Models
- Language Representations Can be What Recommenders Need: Findings and Potentials
- Can LLMs Enhance Performance Prediction for Deep Learning Models?
- GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoning:用 GNN 做检索、增强 LLM 推理;
- 完整版还收录 Circuit Representation Learning with Masked Gate Modeling and Verilog-AIG Alignment(circuitfusion_iclr25)等电路领域工作
此外还有专门的 Prompting 主题:Edge Prompt Tuning for Graph Neural Networks 与 Does Graph Prompt Work? A Data Operation Perspective with Theoretical Analysis,以及 Pre-Training 主题:Holographic Node Representations: Pre-training Task-Agnostic Node Embeddings 与 Centrality-guided Pre-training for Graph。
时序与动态图:从遗忘学习到对抗攻击
Temporal Graphs(时序图)
- N-ForGOT: Towards Not-forgetting and Generalization of Open Temporal Graph Learning
- Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph Forecasting
- TGB-Seq Benchmark: Challenging Temporal GNNs with Complex Sequential Dynamics
- MEMFREEZING: TOWARDS PRACTICAL ADVERSARIAL ATTACKS ON TEMPORAL GRAPH NEURAL NETWORKS
Dynamic Graphs(动态图)
- Rationalizing and Augmenting Dynamic Graph Neural Networks
- Explanations of GNN on Evolving Graphs via Axiomatic Layer edges
完整版清单在此之外还单独列出了 Time Series 分类,说明"时序图+时空预测"在 ICLR 2025 已形成独立研究赛道。
有向图、DAG 与组合优化
Directed Graphs(有向图)
- What Are Good Positional Encodings for Directed Graphs?
- A Benchmark on Directed Graph Representation Learning in Hardware Designs
- Graph Neural Networks for Edge Signals: Orientation Equivariance and Invariance
Directed Acyclic Graphs(DAG)
- LayerDAG: A Layerwise Autoregressive Diffusion Model for Directed Acyclic Graph Generation
- Neural Spacetimes for DAG Representation Learning
- Beyond Directed Acyclic Computation Graph with Cyclic Neural Network
Combinatorial Optimisation(组合优化):完整版清单在此细分出 Linear Programming 与 Mixed-Integer Linear Programming 两个子分类,包括:
- BiQAP: Neural Bi-level Optimization-based Framework for Solving Quadratic Assignment Problems
- Learning to Select Nodes in Branch and Bound with Sufficient Tree Representation
- BTBS-LNS: Binarized-Tightening, Branch and Search on Learning LNS Policies for MIP
科学计算:分子、蛋白质、生物信息学与物理
图表示学习在科学领域的应用在 ICLR 2025 中占据很大比重:
Bioinformatics(生物信息学):空间转录组学、基因网络、药物相互作用预测等 5 篇,包括 Estimation of single-cell and tissue perturbation effect in spatial transcriptomics via Spatial Causal Disentanglement、GeSubNet、DHENN、CBGBench、Global Context-aware Representation Learning for Spatially Resolved Transcriptomics。
Molecules(分子):6 篇,涵盖分子图稀疏表示、构象预测、逆分子设计等,如 A Theoretically-Principled Sparse, Connected, and Rigid Graph Representation of Molecules、REBIND、Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space。完整版还单独设有 Proteins 与 RNA 分类。
Equivariance(等变性):7 篇,代表如 E(3)-equivariant models cannot learn chirality: Field-based molecular generation、VN-EGNN: E(3)- and SE(3)-Equivariant Graph Neural Networks with Virtual Nodes、Geometry-aware RL for Manipulation of Varying Shapes and Deformable Objects。Invariance 主题则包含 Invariant Graphon Networks: Approximation and Cut Distance(完整版对应 iwns_iclr25)。
Physics(物理):PDE 系统的物理编码消息传递等 4 篇,如 PhyMPGN: Physics-encoded Message Passing Graph Network for spatiotemporal PDE systems、SPARK: Physics-Guided Quantitative Augmentation for Dynamical System Modeling。
联邦学习、迁移与持续学习
- Subgraph Federated Learning(子图联邦学习):Subgraph Federated Learning for Local Generalization 与 Decoupled Subgraph Federated Learning;
- Subgraphs for Generalisation:AdaRC: Mitigating Graph Structure Shifts during Test-Time(完整版对应 Matcha,见 adarc_iclr25)、Diversifying Spurious Subgraphs for Graph Out-of-Distribution Generalization;
- Transfer Learning(迁移学习):GraphBridge: Towards Arbitrary Transfer Learning in GNNs 与 Towards Continuous Reuse of Graph Models via Holistic Memory Diversification;
- Diffusion(图扩散) 主题还包含 Instant Policy: In-Context Imitation Learning via Graph Diffusion、Bundle Neural Network for message diffusion on graphs、Relation-Aware Diffusion for Heterogeneous Graphs with Partially Observed Features。
完整分类速查表(255 篇)
下表汇总了完整版 publications_iclr25/README.md 的主题分类结构,供快速导航:
| 分类 | 收录数量(参考) | 代表性方向 |
|---|---|---|
| Scalability | 8 | 图数据集压缩、图稀疏化、可扩展层设计 |
| Heterophily / Oversmoothing / Oversquashing | 8 | 异质图、过平滑、过挤压 |
| Long-range Dependencies / Curvature | 5 | 长程依赖、曲率重连 |
| Spectral Methods / Random Walks | 5 | 谱方法、随机游走 |
| Knowledge Distillation | 2 | 力场蒸馏、符号发现 |
| Hypergraphs / Topological Deep Learning | 9 | 超图、拓扑深度学习 |
| Expressivity(含同态、泛化) | 12 | 表达力理论 |
| Distribution Shifts / OOD / Anomaly Detection | 6 | 分布偏移与异常 |
| Graph Diffusion / Graph Matching | 若干 | 扩散生成、图匹配 |
| Adversarial Attacks / Robustness / Poisoning | 8 | 安全与鲁棒 |
| Explainability / Game Theory | 若干 | 可解释性、博弈 |
| Temporal Graphs / Dynamic Graphs / Time Series | 10+ | 时序与动态图 |
| Knowledge Graphs / Heterogeneous Graphs | 若干 | 知识图谱、异质图 |
| Transfer Learning / Pre-Training / Prompting | 7 | 迁移、预训练、提示 |
| Graph Transformers / Graph Language Models / LLM | 10+ | 图 Transformer、图语言模型 |
| Directed Graphs / DAG | 6 | 有向图、DAG |
| Proteins / RNA / Molecules / Equivariance / Invariance | 20+ | 科学计算 |
| Diffusion Models / Flow-based Modelling / Physics | 若干 | 生成模型与物理 |
| Reinforcement Learning / Robotics / Training | 若干 | 强化学习、训练 |
| Combinatorial Optimisation / LP / MILP | 若干 | 组合优化 |
| Uncertainty / Benchmarks / Miscellaneous | 若干 | 校准、基准、杂项 |
数量为仓库清单的参考统计(255 篇总数),分类与论文条目均可在 publications_iclr25/README.md 中逐一核对。
如何用这份清单:检索、引用与追踪
- 按主题速览:先看 submissions_iclr/2025.md 的 30+ 个主题分组,快速定位与自己的研究方向匹配的论文;
- 取完整清单:需要更全的收录时,使用 publications_iclr25/README.md,其中每个条目都链接到仓库内对应论文子目录;
- 引用论文:进入论文子目录(如 musegnn_iclr25/README.md),复制其中的 BibTeX 条目即可,格式示例:
@inproceedings{musegnn_iclr25,
title = {MuseGNN: Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy},
author = {Haitian Jiang and Renjie Liu and Zengfeng Huang and Yichuan Wang and Xiao Yan and Zhenkun Cai and Minjie Wang and David Wipf},
booktitle = {The Thirteenth International Conference on Learning Representations (ICLR)},
year = {2025}
}
- 读原文与讨论:每个子目录的
links一节都提供 OpenReview 论坛链接,可直达论文全文、审稿意见与作者回复; - 横向对比:由于完整版清单把论文按 60 余个主题细分(如 Curvature、Proteins、MILP 等),你可以轻松对比同一主题下不同论文的方法路线,快速建立"主题-方法-论文"的对应关系。
这份清单的价值在于:它不仅是 ICLR 2025 图学习论文的静态索引,更是一张按主题织成的"研究地图"——从理论(表达力、泛化界)到结构(谱方法、拓扑、超图),从可靠性(鲁棒、投毒、OOD)到应用(分子、蛋白质、物理、时序),再到当下最热的大模型交叉(图语言模型、GNN-RAG、Prompt Tuning),足以支撑一次完整的 ICLR 2025 图学习文献调研。