Showing 36 papers for 2026-06-01
Can Subgraph Explanations Be Weaponized to Steal Graph Neural Networks? introduces a model extraction attack on graph classification under strict black-box conditions, where the attacker only observes discrete class labels and binary explanation masks. The authors show how explanations can be exploited to recover the target model and demonstrate the first such attack in Graph MLaaS settings.
GC-MoE: Graph-Conditioned Mixture of Graph Neural Network Experts for Traffic Forecasting proposes a graph-conditioned mixture-of-experts framework that assigns each road segment a personalized combination of frozen forecasting experts based on graph topology and the recent traffic input window. This node-wise specialization accounts for heterogeneous dynamics across the road network and improves predictive accuracy.
DisasterLex introduces a knowledge-graph–mediated framework that links disaster-domain expert concepts to schemas, enabling geospatial reasoning over heterogeneous hazard, exposure, vulnerability, and lifeline data. By grounding natural-language queries in an expert-driven concept-to-schema KG, it improves cross-schema querying and causal reasoning for disaster response.
Spatio-temporal stochastic graph-based learning for infectious disease forecasting proposes a spatio-temporal stochastic graph-based architecture that injects stochastic formulation into the learning process to produce robust forecasts of infectious diseases over large-scale country data. The model aims to quantify uncertainty and better capture spatial interactions among regions, improving forecast reliability.
AbstainGNN: Teaching Graph Neural Networks to Abstain for Graph Classification introduces AbstainGNN, a theory-guided framework enabling GNNs to abstain when uncertainty is high rather than forcing a prediction. The method provides mechanisms to calibrate abstention and select safe predictions, improving reliability in safety-critical graph classification tasks.
Physics-Informed Coarsening for Multigrid Graph Neural Surrogates presents a multigrid GNN for solid mechanics that combines an encoder–processor–decoder with physics-informed coarsening to preserve elastic and transient behavior. This physics-guided coarsening improves accuracy and stability of surrogate models for deformable solids.
An Efficient and Scalable Graph Condensation with Structure-Preserving introduces SP-ESGC, a decoupled graph condensation method that compresses large graphs into compact, structure-preserving representations. The approach improves efficiency and generalization across different GNN architectures by separating condensation objectives from graph structure preservation.
UniRTL: Unifying Code and Graph for Robust RTL Representation Learning proposes UniRTL to fuse RTL code and graph-based representations (CDFG) into a unified framework, preserving semantic and structural information for more robust RTL representations. This dual-modality approach improves downstream tasks like synthesis, verification, and optimization.
Graph Neural Networks Are Not Continuous Across Graph Resolutions shows that GNNs are not continuous with respect to graph convergence; similar graphs at different resolutions can yield substantially different latent embeddings due to information-propagation schemes. Identifies a structural obstruction in common GNN architectures causing this discontinuity.
Scaling Higher-Order Graph Learning with Maximal Clique Complexes introduces simplified and factored CWL tests (sCWL, fCWL) that keep CWL expressivity while reducing computation, plus the maximal clique complex to enable scalable higher-order GNNs; discusses avoiding explicit clique enumeration.
DG-CoLearn: An Efficient Collaborative Learning Framework for Dynamic Graphs proposes DG-CoLearn, a client-oblivious collaborative framework for dynamic graphs using incremental snapshot processing to reduce retraining cost. It enables cross-partition learning with privacy-preserving data sharing while handling evolving graph structure.
On Efficient Scaling of GNNs via IO-Aware Layers Implementations analyzes that many layers create edge-wise intermediates and memory traffic; classifies into SpMM-based, reductions, and attention-based; offers IO-aware implementation strategies to improve scalability on large graphs.
TRINE: A Token-Aware, Runtime-Adaptive FPGA Inference Engine for Multimodal AI describes TRINE, a single-bitstream FPGA accelerator that runs end-to-end multimodal models without reconfiguration by mapping layers to a unified engine with mode-switching among weight-output-stationary, 1xCS SIMD, and a routed adder tree. Supports heterogeneous architectures.
Graph Machine Learning in the Era of Large Language Models (LLMs) reviews how LLMs intersect with graph ML, discussing opportunities and challenges for integrating GNNs with LLMs, including prompting, reasoning, and knowledge integration.
Aggregation Buffer: Revisiting DropEdge with a New Parameter Block re-examines DropEdge and shows its limited gains due to fundamental architecture limits, and introduces Aggregation Buffer as a new parameterized component to enhance training dynamics and performance.
Beyond Tokens: Enhancing RTL Quality Estimation via Structural Graph Learning proposes incorporating structural semantics from the control data flow graph (CDFG) into RTL quality estimation, complementing LLM-based embeddings to capture structural signals, improving estimation accuracy.
Adaptive Node Feature Selection For Graph Neural Networks proposes a data-, model-, and task-agnostic method to identify and prune irrelevant node features during training, enabling more compact representations and improved interpretability.
View Space: Learning Representation across Arbitrary Graphs introduces view space, a structure-induced representational axis for graphs with heterogeneous data, enabling inductive inference on unseen datasets by unifying disparate numerical features within graph structure.
Scalable Topology-Preserving Graph Coarsening: Concepts and Algorithms proposes STPGC with graph strong collapse and graph edge collapse, providing topology-preserving, scalable coarsening algorithms to maintain GNN performance on coarsened graphs.
Plain Transformers are Surprisingly Powerful Link Predictors argues that plain Transformer architectures can achieve competitive or superior performance on link prediction without heavy structural encodings, offering scalable alternatives to Graph Transformers.
Beyond ReLU reframes oversmoothing in deep GNNs as convergence to a stable homogeneous fixed point via bifurcation theory. The authors argue that the usual stability can be disrupted by replacing standard monotone activations with nonmonotone or topology-informed options, enabling deeper networks with richer representations.
We present L2G-Net, a local to global spectral GNN that factorizes the graph Fourier transform (GFT) into subgraph operators. This exact factorization enables long-range spectral filtering while preserving vertex locality, addressing the cost and locality issues of classical spectral GNNs.
MASPOB introduces a bandit-based prompt optimization framework for multi-agent systems that use graphs and LLMs. It aims to improve MAS performance with sample-efficient prompt search under deployment constraints where workflow changes are limited.
Njord is a probabilistic GNN for ensemble ocean forecasting. It merges a deep latent variable model with graph-based dynamics to produce stochastic forecasts with a single forward pass. It scales to global grids and high-resolution regional domains using scalable graph computations.
BoxLitE maps concepts to convex regions in embedding space to encode hierarchies faithfully. More general concepts correspond to larger boxes, and the model is trained via convex optimization to preserve ontological structure.
Not all toxicity signals are inferable from molecular structure alone, according to an operational taxonomy of explainability gaps in graph-based toxicity prediction. The work delineates the structural information limits that constrain structure-based toxicity modeling.
We propose generating graph-like rules for knowledge graph reasoning with diffusion models. Diffusion-based rule generation captures richer relational patterns beyond chains, including cycles and branches, while mitigating combinatorial search.
GraphARC provides a comprehensive graph-based abstract reasoning benchmark, generalizing ARC to graphs. Tasks require inferring a transformation rule from a few input-output graphs and applying it to a new test graph, covering local, global, and hierarchical transformations.
TraceGraph introduces a graph-based framework to analyze agent trajectories. It constructs a shared decision landscape from pooled rollouts by building a graph over observable states, then highlights productive cores and trap regions to diagnose and summarize each rollout.
HypoAgent presents an agentic framework for interactive abductive hypothesis generation over knowledge graphs. It supports grounding evolving natural language intents across multi-turn dialogues and provides diagnostics when generated hypotheses fail.
Reading Between the Citations introduces a typed claim network for scientific literature where cross-document references are represented as typed claims with source, target, claim text, and a four-class stance. This enriches knowledge graphs with evaluative content about how papers receive and cite each other.
The paper analyzes the linguistic inductive bias of large language models for spatial reasoning in navigation planning. It argues that the linguistic description of spatial representations shapes model behavior and proposes a dual intervention framework to separate topological from geometric influences.
This study investigates masked diffusion models for graph-to-text generation, analyzing generation trajectories. It finds that MDLMs prioritize entities first, then relations and function words, with structural tokens resolved last, and that supervised fine-tuning can disrupt this pattern and cause failures.
PASTA offers a scalable framework for multi-policy AI compliance evaluation. It introduces a model-card style input schema, a policy normalization scheme, and efficient tooling to evaluate AI systems against multiple policies concurrently.
SafeRx-Agent proposes a knowledge grounded multi-agent framework for safe and explainable medication recommendations. It combines grounding of clinical evidence with safety verification and traceability beyond standard drug-code predictions.
Adaptive Prefix-Aware Optimization (APAO) addresses training inference mismatch in generative recommendation by aligning optimization with prefix aware generation. It adjusts the training objective to match how models generate ranked items during inference.