Showing 49 papers for 2026-06-09
Self-healing smart grids can quickly adjust network configuration during outages using reinforcement learning combined with spectral graph neural networks. The approach aims to automate reconfiguration (switching, load shedding) faster and with less compute than traditional ML, reducing power disruptions.
We propose a topological framework to compare trained GNNs by embedding SBM-induced structures on the graphon-signal space of an MPNN onto the unit sphere. The method leverages cut-distance compactness and Frieze–Kannan regularity to analyze GNN representations.
The paper develops a GNN framework to classify finite groups as solvable or non-solvable using graph representations, including Cayley graphs. It tests on unseen groups to assess whether GNNs can learn underlying algebraic solvability properties beyond training data.
Introduces KITE, a tri-modal transformer that jointly models text, images, and factual knowledge graphs for fake news detection. It goes beyond text-image fusion and treats external knowledge as an integral modality to detect semantic inconsistencies.
Proposes ST-GNN with a learnable Tweedie distribution head to handle sparse vessel traffic data characterized by bursts and zeros. The model can forecast non-zero activity more robustly than standard ST-GNNs and addresses zero-inflation with a learnable head.
Shows GNNs for DDI prediction typically rely on molecular graphs; augment with pharmacogenomic knowledge from PharmGKB to provide metabolic pathway context, mitigating the information ceiling of labels. Experiments indicate improved predictive performance.
EssentialGIN introduces a graph-isomorphism-based neural approach for gene essentiality prediction, aiming to overcome limitations of simple centrality metrics. It seeks to improve accuracy for identifying essential genes with computational methods.
Examines graph reconstruction attacks (GRA) on GNNs, showing when and why adjacency information can be recovered from features, labels, embeddings, or predictions. Also discusses defensive strategies against such attacks.
Presents Mesh Graph Network (MGN) for predicting von Mises stress on 2D components with arbitrary hole geometries, enabling FE-like predictions across varying geometries. It aims to generalize beyond fixed meshes to accelerate simulations.
QueryWeaver converts natural language queries into structured graphs and uses a deterministic DFS planner to orchestrate multi-tool executions, improving reliability for cross-tool queries.
GeoGNN proposes a time-series geolocalization approach with a two-tower GNN: the spatial tower leverages a geographic adjacency graph to embed location candidates, enabling inference of the time series' geographic origin.
Argues for inductive cross-network generalisation for Graph Foundation Models in dynamics of complex networks, and outlines four design properties. ts-net trained on synthetic multilayer networks demonstrates cross-network generalization.
Investigates calibration attacks on GNNs, highlighting challenges in graph calibration under adversarial perturbations; discusses how discrete graph structure complicates optimization and the limitations of underconfidence objectives.
Presents adaptive loss balancing to improve noise-robust GRPO in generative recommendation; analyzes biased reward signal from production rankers and proposes balancing strategies to stabilize RL-guided generation.
Offers a generalized rank-based evaluation framework for knowledge graph completion, discussing perspectives and analyses beyond conventional MRR/Hits@K metrics to better reflect real-world use cases.
Demo of PROBE-Web, an interactive system to probe evaluation landscapes of KGC models by adjusting predictive sharpness and popularity-bias robustness, enabling flexible model comparison.
Heterophily-aware adaptive knowledge distillation (HADES) improves hypergraph neural networks by tailoring distillation to heterophilic nodes, addressing reliability gaps of teacher guidance.
Beyond Convolution: Advancing Hypergraph Neural Networks with Hypergraph U-Nets
OnlyDense introduces reduced-order modeling for Lagrangian simulations like SPH/MPM, aiming to reduce computational cost while capturing key dynamic features across scales.
Shows the Injection Paradox in safety-trained RAG-based LLM recommendations: prompt injections in retrieved documents can backfire, suppressing the target brand’s recommendations not only in the injected doc but across the same brand’s other docs.
PRISM introduces a topology-aware imputation for modality-deficient federated graph learning. It addresses clients lacking a modality by leveraging graph topology to reconstruct missing descriptions or images, enabling cross-modal collaboration without requiring shared modalities. The method demonstrates improved performance under modality deficiency.
Graph Mamba Operator (GraMO) is a latent-simulator for interacting particle systems. It captures spatial interactions and long-range temporal dependencies with a latent operator to reduce error accumulation in long horizons and to model multi-hop and global structure, addressing limitations of autoregressive and local GNNs.
Efficient Traffic Prediction at Scale systematically studies the STGCN depth. It evaluates whether current STGNN architectures are over-parameterized and provides insights into depth vs. efficiency trade-offs for scalable ITS deployments.
CAPruner proposes a Conceptual-Adjacent Scene Graph Pruner to improve 3D spatial reasoning in LLMs. By pruning less task-relevant, conceptual-adjacent relations in scene graphs, it reduces token costs while preserving essential spatial relations.
ALCMeans proposes Automatic Laplacian Centrality Means for unsupervised community detection. It uses local Laplacian energy to automatically determine the number of centers and improve scalability and accuracy over traditional methods.
Graph-to-SFILES uses graphs to predict control structures from process topologies. It treats flowsheets as graphs and generates SFILES representations with a graph-based generative AI model, enabling permutation-invariant control structure prediction.
TaskPGM learns task mixtures via a probabilistic graphical model over tasks. It uses an energy-based model on a Markov random field to capture task affinities and automatically allocate training budget for supervised fine-tuning.
GraCE-VAE proposes causal representation learning from network data. It uses observed relational context and a graph-aware VAE to disentangle causal factors under linear interventional faithfulness, leveraging known networks like protein interactions.
A Graphop Analysis of GNNs on Sparse Graphs provides a unified framework by defining a compact metric on graph spaces, enabling generalization and universal approximation results for MPNNs on both dense and sparse graphs.
CTS-Bench benchmarks graph coarsening trade-offs for GNNs in Clock Tree Synthesis in EDA; provides benchmark suite to evaluate memory/time vs accuracy.
Graph-GRPO trains Graph Flow Models with reinforcement learning under verifiable rewards. It aligns GFMs with task/objective rewards online, enabling better optimization with human preferences.
GraphER enhances RAG by graph-based enrichment and reranking to ensure complete evidence across multiple sources, addressing inefficiencies of iterative agentic retrieval.
Capacity-Controlled Global Attention for Graph Transformers argues that global softmax attention conserves mass per head, causing over-smoothing, low-rank bottlenecks, and other pathologies; proposes a capacity-controlled mechanism to alleviate.
A Survey of Heterogeneous Graph Neural Networks for Cybersecurity Anomaly Detection reviews HGNN methods for anomaly detection in cybersecurity, highlighting heterogeneity and temporal dynamics.
Towards Personalized Bangla Book Recommendation: A Large-Scale Heterogeneous Book Graph Dataset introduces RokomariBG, a large-scale heterogeneous Bangla book graph dataset for personalized recommendations.
What Makes a Desired Graph for Relational Deep Learning? studies how schema-derived graphs may hurt relational reasoning; proposes controlled structural adaptation to produce graphs better suited for GNNs.
Extending Ontologies: From Dense Embeddings to Hybrid Quantum-Fuzzy Systems surveys integrating ontologies and knowledge graphs with dense embeddings; advocates hybrid quantum-fuzzy representations.
ZIPP: Zero-shot Image Personalization from Personas enables per-user style personalization for image generation without dense interaction histories or fine-tuning, addressing cold-start.
Optical Reasoning: Rethinking Images as an Expressive Reasoning Medium Beyond Text proposes using images as the reasoning medium for both language and multimodal tasks, i.e., optical reasoning.
Implicit Causal Graph Construction in Text via Chain Discovery studies constructing causal graphs from text by treating described cause-effect pairs as endpoints of latent graphs; uses LLMs to infer intermediate events; compares end-to-end graph construction vs chain discovery approaches.
This work demonstrates a vulnerability where payloads in plain English can bypass detectors when they are encoded as structured float parameters and reconstructed only as fragmented telemetry. Across 14,400 attack trials on three commercial LLM APIs, the float-array carrier achieves a high evasion rate (about 94.3%), revealing a steganographic pathway for indirect prompt and content injection.
This report summarizes the CHIIR 2026 Workshop on Generative AI and Academic Search (GAI&AS) and how GenAI is reshaping academic search systems. It discusses challenges and opportunities in designing and evaluating future systems that integrate GenAI for summarization, recommendation, synthesis, and conversational interaction beyond traditional document retrieval.
We introduce Fast Adaptive Semantic Entropy (FASE), a fast metric that approximates functional correctness in multi-agent code generation without relying on costly LLM-driven equivalence checks. By efficiently estimating uncertainty, FASE reduces the need for expensive verification and guides cooperative agent behavior.
We propose a diachronic modeling pattern for legal norms using the LRMoo ontology, enabling component-level versioning and deterministic point-in-time reconstruction of legal texts. The approach models evolution as a chain of versioned F1 Works and distinguishes language-agnostic Temporal Versions to support granular, event-centric knowledge graphs of legal norms.
We evaluate AI-driven formal proof search in mathematics using Lean; the agent autonomously solved 9 of 353 open Erdős problems at a moderate cost, and proved 44 of 492 OEIS conjectures. The results demonstrate potential for AI-assisted proof in several areas of combinatorics, optimization, graph theory, and beyond.
AgroOmni provides a large-scale, multi-view agricultural dataset spanning ground-level, UAV, and satellite imagery to support cross-scale multimodal reasoning. The dataset addresses ground-scale bias and scale-confusion that cause semantic collapse in current multimodal language models, enabling more robust cross-scale agricultural understanding.
OneFeed proposes a unified generative framework that connects feed content enhancement and query generation, addressing the separation between feed-driven implicit signals and explicit user queries. It enables feedback between feed interactions and generated queries to improve candidate retrieval and query quality.
HARPO introduces Hierarchical Agentic Reasoning with Preference Optimization for user-aligned conversational recommendations. It addresses the shortcoming of proxy metrics that optimize retrieval or fluency rather than true recommendation quality, by using hierarchical reasoning to better align with user preferences under uncertainty.
DP4SQL presents a differentially private SQL system with flexible privacy policies for relational databases. It goes beyond rigid, single-policy DP, enabling protection configurations across multiple relations and allowing customization of which data pieces receive plausible deniability guarantees, including existence protection.