Showing 53 papers for 2026-03-03
They develop a property-driven evaluation framework to assess GNN expressiveness at scale, grounded in formal specifications using Alloy. The study proposes a configurable graph dataset and systematic empirical evaluation to understand which graph properties GNNs can capture.
They propose Dynamic Spatio-Temporal Graph Neural Network (DST-GNN) for early detection of pornography addiction using EEG. It combines Phase Lag Index based Graph Attention Network for spatial connectivity with a bidirectional gated recurrent component for temporal dynamics.
The paper introduces a policy-guided approach to synthesize graph outliers for unsupervised OOD detection, replacing fixed sampling heuristics with a learned policy that explores informative OOD regions. This yields improved OOD detection boundaries and robustness.
We adopt a novel Riemannian-geometry perspective to multi-domain graph pre-training and building graph foundation models. The method questions how knowledge is integrated and transferred across domains, proposing a new way to align and transfer representations.
GeMi is a graph-based multimodal recommender for narrative scroll paintings. It leverages graph representations to capture relationships among items and users across modalities (images, text, etc.) to improve recommendations.
Fed-GAME is a personalized federated learning framework for time-series forecasting that uses a learnable dynamic implicit graph to model client relationships. It uses a decoupled parameter-difference update protocol on the server, decomposing updates into streams for personalization and global aggregation.
Invariant-Stratified Propagation enhances expressivity by distinguishing nodes based on their structural roles and propagating messages accordingly. This approach targets the expressivity limits of standard message-passing (1-WL) while avoiding prohibitive computation costs.
The work studies OOD generalization in dynamic graphs via causal invariant learning, aiming to identify invariant and variant evolution patterns and derive evolution rationales. The goal is to improve generalization across diverse OOD shifts with limited data.
FreeGNN is a continual source-free graph domain adaptation framework for renewable energy forecasting, enabling adaptive predictions at unseen sites without source data or target labels. It relies on a spatio-temporal graph modeling to align domain distributions.
BAED introduces a new paradigm for few-shot graph learning with an explanations-in-the-loop component, guiding learning with explanations to improve adaptation to new label distributions.
The paper mitigates topology biases in graph diffusion models via counterfactual intervention on topology, reducing bias from sensitive attributes. This approach enables fairer synthetic graphs and more trustworthy diffusion-based generation.
Proposes Graph Concept Bottleneck Layer to reveal and quantify combinatorial reasoning in GNNs; it exposes how topological concepts contribute to predictions and quantifies their influence, moving beyond post-hoc explanations.
A knowledge-graph based framework unifies heterogeneous evidence sources (ChEMBL, PubMed, ClinicalTrials.gov, FAERS) into an evidence-weighted graph to study safety of protein kinase inhibitors.
GrapHist is a graph-based self-supervised learning framework for histopathology that models tissues as cell graphs to learn biologically informed representations capturing cell interactions.
GCL-Sampler proposes a graph-contrastive learning approach to automatically discover kernel similarity from GPU trace graphs, enabling more effective sampling for faster GPU simulations.
Heterophily-Agnostic Hypergraph Neural Networks with a Riemannian Local Exchanger address heterophily in high-order relationships by a Riemannian exchange mechanism for robust message passing.
The work develops graph neural network-based super-resolution for turbulent reacting flows on complex meshes, reconstructing small-scale structures from coarse simulations on unstructured meshes.
Graph neural network force fields for adiabatic dynamics of lattice Hamiltonians enforce lattice translation and point-group symmetries through local message passing, enabling scalable, quantum-accurate simulations.
FSW-GNN presents a Bi-Lipschitz WL-Equivalent GNN that improves graph separability on WL-equivalent graphs by using a Lipschitz-controllable transformation, addressing limitations of standard WL-equivalent networks.
UFGraphFR proposes a graph federation recommender using user text description features to derive semantic similarity among users, enabling privacy-preserving cross-user recommendations.
We show that Graph Neural Networks and Transformers can learn to reason about geometric constraints. By training on predicting the spatial position of points on a discrete 2D grid from constraint descriptions, both models not only predict positions but also form the hidden figures in embedding space during reasoning, effectively recovering the grid structure.
We extend topological deep learning to directed semi-simplicial structures to capture higher-order directed patterns in data such as brain activity. Unlike existing undirected TDL models, our approach represents multi-way relationships with directionality and demonstrates improved brain activity decoding.
We propose rapid training of Hamiltonian graph networks by using random features to accelerate optimization, reducing training time while preserving physics-informed permutation invariance. Benchmarking against common optimizers shows substantial speedups for large, complex systems.
Introducing GraphUniverse, a framework to generate whole families of synthetic graphs, enabling systematic evaluation of inductive generalization at scale. It creates graphs with persistent structural properties, allowing controlled studies of how models generalize to unseen graph families.
We propose LEAP, a local, learnable graph positional encoding based on the Euler Characteristic Transform (ECT). The differentiable approximation of ECT provides topology-aware encodings that augment MPNNs, improving performance on geometry- and topology-sensitive tasks.
We analyze failures of Self-explainable GNN explanations (SE-GNNs) and show that explanations can be unrelated to how the model infers labels. We propose diagnostics to identify such misleading explanations and discuss implications for trust and safety.
We introduce EDT-Former, an Entropy-guided Dynamic Token Transformer that aligns graph representations with LLMs for molecular understanding. By allocating dynamic tokens based on information content, the approach avoids fixed tokens and extra fine-tuning, while capturing stereochemistry and substructural context.
We present MSH-LLM, a Multi-Scale Hypergraph method that aligns LLMs for time-series analysis. By constructing multi-scale hypergraphs and a hyperedge mechanism, the approach captures complex temporal dependencies and enables improved reasoning when processing time-series data with LLM backbones.
We develop theoretical foundations for Superhypergraphs and Plithogenic GNNs, extending hypergraph modeling to nested, set-valued relations. The work analyzes properties and learning implications, setting the stage for new GNN architectures capable of hierarchical, higher-order interactions.
HGTS-Former introduces a Hierarchical HyperGraph Transformer for multivariate time series. It models complex variable couplings with hypergraphs and hierarchical attention, improving performance on challenging tasks.
We present action-conditional GNNs for robotic peg insertion, extending FIGNet with new node and edge types to predict motion and force-torque during contact. The model learns in a self-supervised way using only joint encoder and force-torque data while touching.
DAG-Math models chain-of-thought as a rule-based stochastic process on directed acyclic graphs, where nodes are derivation states and edges encode rule applications. We define logical closeness as a metric to quantify proximity between derivation states.
Nazrin introduces atomic tactics—a small finite set of tactics capable of proving any provable statement in Lean 4—and a transposing atomization algorithm that rewrites arbitrary proof expressions into a sequence of atomic steps.
MED-COPILOT combines guideline-grounded GraphRAG retrieval with hybrid semantic-keyword similar-patient retrieval to support clinicians with evidence-sourced decision making in an interactive clinical decision-support system.
MMCOMET is a large-scale multimodal commonsense KG that extends ATOMIC2020 with visual information, yielding over 900k multimodal triples and enabling multimodal reasoning for tasks like image captioning and storytelling.
FCN-LLM enables LLMs to understand brain functional connectivity networks via graph-level multi-task instruction tuning, aligning FCNs with textual modalities and enabling reasoning over connectivity graphs.
GraphScout equips LLMs with intrinsic exploration ability for agentic graph reasoning, enabling iterative, graph-based reasoning with retrieval augmentation beyond fixed prompts.
We construct a pharmacology knowledge graph from ChEMBL 36 and perform strict temporal validation to quantify the value of chemical structure for drug repurposing; benchmarking multiple KG models with hard negatives shows insights into when structure matters.
We present Self-Healing Router, a fault-tolerant tool-routing architecture for tool-using LLM agents. It uses parallel health monitors and priority-based routing to tolerate tool outages and failures, balancing reliability, latency and cost.
We propose a joint sensor deployment and physics-informed graph transformer detector, optimizing sensor locations with NSGA-II while training a PIGTN-based detector in a closed loop, achieving improved attack detection under practical constraints.
Proposes a graph-theoretic agreement framework to verify and secure distributed multi-agent LLM systems. It analyzes modern coordination patterns such as multi-agent debate, constitutional oversight, and helper-critic loops that rely on adversarial critique for error correction and reasoning refinement. Because LLMs are dynamic systems with latent states that are not fully observable from their outputs, the framework aims to formalize coordination and security properties in these networks.
Introduces SSKG Hub, an expert-guided platform that converts sustainability standards (e.g., GRI, SASB, TCFD, IFRS S2) into auditable knowledge graphs via an LLM-centered pipeline. The system supports automatic standard identification, configurable chunking, standard-specific prompting, and robust auditing and provenance tracking to support downstream analysis.
KG-Followup introduces a knowledge-graph–augmented LLM with active in-context learning to generate relevant and important follow-up questions for pre-diagnostic assessment. A structured medical knowledge graph guides question selection and ensures clinical relevance, serving as a patch to the diagnostic workflow. This module aims to improve pre-diagnostic interactions and decision support.
Graph-Tokenizing LLMs (GTokenLLMs) encode graph structures and long texts into a graph token sequence, enabling graph-based input to LLMs. The model is then aligned with natural language tokens via reconstructive graph instruction tuning to improve graph understanding and reasoning.
G-reasoner presents foundation-model–based unified reasoning over graph-structured knowledge, addressing the limitations of static parametric knowledge. It combines retrieval-augmented generation with graph-aware modeling to better reason about graph data, building on GraphRAG ideas. The goal is more coherent, knowledge-structure–aware reasoning in LLMs.
KG-guided chain-of-thought (CoT) for visit-level disease prediction on MIMIC-III maps ICD-9 codes to PrimeKG and mines disease-relevant nodes and paths. These paths scaffold temporally consistent CoT rationales, and only samples whose conclusions match observed outcomes are retained; the approach uses lightweight instruction-tuned LLMs (LLaMA-3.1-Instruct-8B).
Hierarchical Multi-Scale Knowledge-aware Graph Network (HMKGN) models multi-scale interactions and spatially hierarchical relationships within whole-slide images for cancer prognosis. It enforces a hierarchical structure with spatial locality constraints, using local cellular-level dynamic graphs to aggregate proximate patches. This design improves prognostic performance over conventional MIL methods.
Reasoning by Exploration presents a unified approach to retrieval and generation over graphs by treating reasoning as exploration, rather than a fixed two-phase pipeline. It addresses the difficulty of faithfully incorporating graph structure in large graphs when using traditional RAG, and aims to improve fidelity and efficiency of graph-based reasoning.
SynthKG is a multi-step data synthesis pipeline that generates high-quality document-level knowledge-graph data to enable scalable KG construction. It reduces reliance on expensive LLMs and uses distillation to produce training data for doc-level KGs, enabling high-quality graphs for large corpora.
Disk-Resident Graph ANNS Search provides a comprehensive experimental evaluation of disk-based graph-based approximate nearest neighbor methods. It analyzes fundamental performance trade-offs across storage layout, memory hierarchy, and execution paradigms to guide design choices for disk-resident ANNS.
CatapultDB accelerates vector search by exploiting query locality in graph-based indices. It makes the traversal workload adaptive to the query stream, reducing redundant hops and revisiting entry points, thereby improving throughput on large-scale vector data.
BAMG proposes a Block-Aware Monotonic Graph Index for disk-based ANNS that co-optimizes storage layout with graph structure. By aligning block layout with monotonic graph traversal, BAMG improves I/O efficiency and query latency for large-scale vector search.
S3GND introduces a learning-based method for subgraph similarity search under generalized neighbor difference semantics (GND). It accounts for keyword-set relationships between vertices to define a more flexible similarity measure, enabling more effective retrieval of semantically related subgraphs in large graphs.