Showing 26 papers for 2026-07-22
This work proposes augmented Graph Neural ODEs with exogenous controls to predict unsteady airfoil aerodynamics across space and time, tackling limitations of conventional CFD and linear models. By integrating exogenous inputs such as gusts into a graph-based dynamic model, it aims to provide accurate, efficient predictions for design and certification.
ChemHyperMag introduces a physics-informed learning framework that uses a functional-group hypergraph to capture asymmetric interactions and motif-level effects in molecules, improving multitask ADMET prediction with missing labels. It builds hypergraphs from rings, BRICS fragments, Bemis-Murcko scaffolds, and bonds, enabling richer relational modeling and potentially driving physical priors.
We propose Scalable and Efficient Joint Scoring Spiking Embedding Architecture for large-scale dynamic graphs, enabling scalable dynamic graph representation without heavy edge-level reconstructions. The method leverages spike-based embeddings and a joint cost-prediction and ranking objective to handle scarce labels and large graphs efficiently.
GNNAS-TSP introduces a Graph Neural Network-based automated algorithm selection framework for the Traveling Salesman Problem. It learns instance representations directly from raw graph data and treats AS as a joint cost prediction and ranking task, analyzing performance under different budget regimes.
GEqTrain provides a configuration-driven framework to retarget equivariant GNNs across 3D scientific tasks. It decouples dataset semantics, model composition, and training objectives with Hydra-based configurations, mapping data to typed fields and assembling model stacks and losses declaratively, enabling reuse across tasks.
We enhance Neural Markov Logic Networks with parallel-noising by combining richer graph neural potentials and a parallel-tempering-inspired training/inference algorithm, enabling better scalability and expressiveness on larger relational structures than standard NMLNs. The approach also seeks to maintain inference efficiency on complex relational graphs.
One Model, Many Graphs investigates using vision-language backbones to unify graph learning across modalities. It addresses modality-heterogeneous attributed graphs (text-only, image-only, or both) and proposes methods to operate with a single model across settings, improving scalability and flexibility.
Toward Auditable Fraud Detection presents a layered, auditable pipeline on PaySim that combines a gradient-boosting classifier, graph-derived features, autoencoder anomaly signals, TreeSHAP explanations, and a bounded LLM agent to investigate uncertain cases, yielding interpretable and reviewable fraud detection. The study also emphasizes removing simulator-specific shortcuts to ensure fair evaluation.
GUIDED Network-Agnostic Feature Initialization proposes a network-agnostic initialization layer to improve spatial transferability of GNNs used for traffic assignment, decoupling feature initializations from fixed network topology and enabling transfer to new urban networks.
We train a GNN to approximate the NP-hard SPR distance between phylogenetic trees, enabling near-constant-time distance estimation after training. The paper provides a dataset of 864 trees across four bacteria and demonstrates the model's ability to generalize across tree shapes and species.
Attacking Graph Foundation Models Through Their Shared Representation shows that the alignment layer of graph foundation models is a new attack surface. The authors perform inference-time perturbations in the shared representation space across multiple public models, causing a collapse in the representation and degrading performance.
GQD-AdsNet uses graph neural networks to rapidly explore adsorption energies of transition metals on graphene quantum dots, enabling faster screening of single-atom catalysts by predicting binding energies for various metal-QD configurations.
Chi-Square Wavelet GNNs for HIN Anomaly Detection introduces ChiGAD, a spectral GNN framework for heterogeneous information networks that captures multi-path semantics, preserves high-frequency information, and handles class imbalance for robust anomaly detection.
Finite-Agent Stochastic Differential Games on Large Graphs II presents graph-based architectures with Non-Trainable Modification (NTM) that sparsify feedforward nets according to the graph, embedding fixed components aligned with topology to compute Nash equilibria in SDGs on graphs.
Budgeted Indirect Adversarial Attack on Graph-Based Anomaly Detection presents BETA, an indirect evasion attack where only a subset of non-target sensors can be perturbed under a budget, aiming to suppress true anomalies or trigger false alarms, guided by graph explanations and centrality.
CLT-Forge is a scalable library for Cross-Layer Transcoders and Attribution Graphs, enabling mechanistic interpretability of LLMs by representing computations as sparse, shared feature graphs across layers and coupling transcoders to produce attribution graphs.
Node-as-Agent proposes Graph Agentic Network, where nodes act as agents to balance informativeness and leverage global semantic relationships beyond local aggregation, enabling adaptive message passing guided by node-level agency. This framework aims to improve performance on heterogeneous graphs by dynamic, agent-like routing of information.
Position argues that AI/ML deepfake research is misaligned with the real-world problem of AI-generated non-consensual intimate imagery (AIG-NCII), highlighting a landscape where technical deepfake interventions largely ignore this abused domain and proposing alignment with real harms.
TRACER presents a trajectory-aware clinical risk prediction framework using severity-grounded knowledge graphs and retrieval-augmented generation to incorporate external knowledge and clinical notes, improving risk prediction under data sparsity and clinical progression dynamics.
PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents
OntoBook converts medical ontology structure into pretraining data for encoder language models. It uses random walks through ontology graphs to capture hierarchical and causal relations between medical codes, then a large language model reformulates these walks into fluent textbook-style prose. The resulting text is used to train ModernCamemBERT, a 149M-parameter French encoder, with joint objectives of masked language modeling and relation prediction between code pairs.
OpenRTAG presents a comprehensive benchmark for robust learning on text-attributed graphs (TAGs) under data quality degradation. It identifies nine degradation scenarios spanning text, structure, and label quality, such as sparsity, noise, and imbalance, and evaluates how these affect TAG learning across tasks and datasets. The study surveys existing mitigation strategies and highlights gaps in evidence.
The paper advocates using multi-relational GCNs (MR-GCNs) for sequential user modeling on graph-structured data. It shows that treating different relation types as homogeneous reduces expressive power, and proposes MR-GCN-based architectures to capture richer semantics across diverse relations for improved downstream predictions in personalized systems.
AgentTrails presents a prototype system for agent provenance to improve trust and reuse in agentic tasks. It argues that trajectories logged by LLM-powered agents conceal the dataflow dependencies between actions and created artifacts, hindering debugging and reuse. The system exposes these dataflow relations to enable better understanding, debugging, and reuse of computations.
SR-Agent introduces an experience-driven framework for post-ranking strategy refinement in e-commerce recommendations. It addresses the issue that static post-ranking rules become stale as the environment evolves and manual updates are costly, by enabling automatic refinement guided by experience data to maintain diversity, similarity balance, and exposure.
GSPRec proposes improving item representations for graph signal processing-based collaborative filtering. It observes that standard graph-based CF acts as a low-pass filter that loses intermediate-frequency components where community-level preferences lie, and that prior work relies on designed filters using only the user-item interaction matrix. GSPRec derives richer item embeddings by incorporating graph spectral signals and sequence-level proximity information beyond the raw interaction matrix to better capture community structure for recommendations.