Showing 27 papers for 2026-05-13
This paper addresses heterophilous graphs where adjacent nodes carry different labels, and existing spectral GNNs struggle with hub dominated aggregation as well as oversmoothing and oversquashing caused by suboptimal polynomial filters. They propose Hierarchical Multi-Scale Graph Neural Networks to mitigate these issues with scalable, degree aware, multi-scale filtering and pooling that better capture distant heterophilous signals.
Protein language models like ESM-2 learn residue representations that are difficult to interpret in structural terms. The authors propose a plug-and-play framework that projects these representations onto protein contact graphs and applies SoftBlobGIN, a differentiable substructure pooling GNN, to perform structure-aware message passing and enhance interpretability.
Knowledge graph completion benefits from region-based relation embeddings but suffers from absolute boundary constraints during optimization. CORE introduces cyclic orthotope relation embeddings to model relations as geometric regions, mitigating boundary issues and improving reasoning over complex patterns.
Neural Sheaf Diffusion generalizes diffusion-based GNNs by using sheaf Laplacians with learned restriction maps to yield task-adapted geometry. The authors provide a quiver-theoretic interpretation by linking cellular sheaves on graphs to representations of the incidence quiver, showing oversmoothing as a degeneracy in the harmonic representation space.
Large language models orchestrate tool use via graphs that encode dependencies, but existing strategies rely on a matching paradigm that often misaligns tool choices with subtask intent. GRAFT proposes Graph-Tokenized LLMs that represent tools as graph tokens to better plan and coordinate multi-step tool usage.
Multimodal Federated Graph Learning suffers semantic drift because local modality encoders do not share a common semantic space. STAGE tackles this drift by aligning semantic spaces across clients and modalities, enabling stable collaboration and robust aggregation in MM-FGL.
We formulate subgraph importance estimation for pretrained GNNs as a linear Group Lasso problem in embedding space, incorporating structural priors about graph substructures. The method is model-agnostic and does not require target labels; experiments show improvements over baselines on real-world graphs.
A unified Graph Language Model for multi-domain and multi-task graph alignment and instruction tuning is proposed. The approach uses GNN encoders and aligns their outputs with LLM tokens through alignment-focused instruction tuning to produce a unified graph token space across domains and tasks.
Message passing is reframed as a linearized graph sequence model, connecting graph learning with advances in sequence modeling. The framework decouples computational steps and recasts graph computations as sequence processing, enabling simpler architectures and better integration with sequence models.
Popularity bias in GNN based collaborative filtering arises from skewed interactions and repeated high order propagation that amplify popular items. The paper proposes a debiasing of message passing to reduce this bias, improving long tail recommendations through reweighting, regularization, and causal techniques.
Full vehicle crash simulations are computationally expensive, motivating learned hybrid surrogates. The paper investigates mesh-based GNN surrogates (MeshTransolver, MeshGeoTransolver, MeshGeoFLARE) to predict time-resolved deformation fields, demonstrating accuracy, spatial regularity, and physics-aware attention that supports engineering interpretation.
Uncertainty quantification in GNNs is essential due to noisy data and missing edges. Random-Set Graph Neural Networks model uncertainty with random-set representations to capture both aleatoric and epistemic uncertainties, yielding calibrated predictions.
SEMIR addresses segmentation of small and sparse structures in large scale images by decoupling inference from dense voxel level computation and by learning representations that emphasize semantic minors, improving segmentation performance and efficiency.
Model level explanations for GNNs are challenging; this work reconstructs logits from a rule to graph readout, where grounded subgraph concepts form logical rules and their embeddings are used to reconstruct the GNN logits, providing global interpretability.
Dynamic long term memory is a bottleneck for language agents. The paper analyzes smoothness errors in dynamics models and unitary convolutions, and proposes strategies to avoid excessive smoothing while preserving stability in flow-based or diffusion-like dynamics.
Personalized storefronts in large marketplaces are built from independent components like placement, retrieval, and ranking. The authors propose a cascaded generative approach that jointly optimizes these components to deliver coherent, personalized, and dynamic storefront experiences.
Multimodal Graph Learning can be made efficient by decoupling propagation and aggregation. CAMPA shows that decoupled pipelines are much more scalable, but modal conflict arises in both stages; the paper analyzes and addresses modal conflict to achieve aligned, efficient learning.
Generative POI recommendation with LLMs can be outpaced by real world changes. AWARE employs an LLM agent to generate world knowledge about events and trends to inform next location recommendations, improving responsiveness to evolving conditions.
Knowledge graph enhanced LLMs with soft prompts introduce a dual channel interface. The paper shows that this architecture creates robustness gaps for backdoor attacks on the graph conditioned channel, proposing BadSKP style attacks and discussing defenses.
SAGE is a self-evolving agentic graph memory engine that treats graph memory as a dynamic long term substrate for structure aware associative memory. It couples a memory writer that incrementally constructs structured memories with downstream reasoning components for improved agent memory.
We propose a computationally efficient hybrid framework that merges citation topology with LLM-based text similarity to reconnect fragmented citation networks. The approach is applied to 662,369 Web of Science publications in Mathematics and Operations Research & Management Science. It augments the original graph by adding semantic edges between small disconnected components and by weighting existing citations to improve connectivity and downstream analyses.
We introduce BifrostRAG, a dual-graph retrieval-augmented generation system for multi-hop question answering in construction safety, modeling both linguistic relationships and document structure. The architecture supports a hybrid retrieval mechanism that combines semantic text similarity with document topology to fetch relevant clauses and support answer generation. This enables robust reasoning across interlinked regulatory clauses for safety compliance.
We study internal mechanisms of LLMs and find that they spontaneously reconstruct a graph's topology, evidenced by a distinct sawtooth pattern in attention maps. We name this phenomenon SLASH the Sink and propose a method to sharpen the internal structural attention, improving graph-structured input understanding without external adapters or fine-tuning. The goal is to enable more cost-efficient processing of graph topologies by LLMs.
TOHA is a topology-based hallucination detector in the RAG setting that uses a topological divergence metric on attention-induced graphs. Examining prompt–response subgraphs reveals that higher divergence in certain attention heads correlates with hallucinated outputs. This provides a robust signal for detecting hallucinations in generated content.
Proposes project-level C-to-Rust translation using Pointer Knowledge Graphs to model pointer relationships and cross-module dependencies across the codebase. Unlike bottom-up, per-function translation, the approach uses knowledge graphs to coordinate translations and enforce safety at the project level. LLM-based code methods can produce safer, more idiomatic Rust code with this guidance.
EndoVGGT introduces a geometry-centric framework for surgical 3D reconstruction with a Deformation-aware Graph Attention (DeGAT) module. DeGAT dynamically constructs feature-space semantic graphs rather than fixed spatial neighborhoods to capture long-range dependencies in deformable tissues and overcome low texture, specular highlights, and occlusions. This improves depth estimation and 3D reconstruction quality in surgical settings.
Graph-Grounded Optimization envisions decision variables, constraints, and objective coefficients sourced from a property knowledge graph via Cypher queries, rather than free-form text or static inputs. It surveys recent LLM/SLM-driven optimization systems and argues for putting property graphs at the center of input modalities. The authors instantiate the paradigm with Rao-Family metaheuristics, classical OR, and SLM-driven formulations to demonstrate viability.