Showing 28 papers for 2026-06-05
This paper introduces In-Context Graphical Inference (ICG-I), an autoregressive Graph Transformer designed to preserve the sequential elimination structure used in exact marginal inference. It argues that the traditional exactness-vs-scalability trade-off arises from a misaligned inductive bias and proposes ICG-I to unify accuracy with scalable inference in discrete graphical models. The approach targets marginal inference on high-treewidth graphs where iterative methods struggle on frustrated topologies.
Graph Cascades introduces a mesoscopic rewiring strategy for Graph Neural Networks and Graph Transformers. Using contagion-based diffusion, it constructs an auxiliary graph where node pairs reinforced by multi-hop signals are promoted to direct neighbors, enabling better capture of intermediate-scale structure in linear time. The authors also provide theoretical conditions under which reinforcement-based rewiring improves performance.
The Graph Set Transformer (GST) learns on sets of graphs by interleaving node-level feature propagation with cross-graph contextual modelling at every layer. Unlike prior approaches that require pre-encoded graph embeddings, GST fuses local graph structure and set-wide context continuously to enable joint reasoning over a collection of graphs.
SpliceBind presents a graph neural network framework for isoform-aware prediction of binding-pocket druggability, enabling comparisons across splice variants. It improves predictive accuracy (AUROC) relative to a state-of-the-art baseline and analyzes when structural methods succeed or fail across six clinically validated variants.
DegradoMap introduces a graph neural network that predicts PROTAC-mediated degradability from protein structure and E3 ligase identity alone, without requiring the full PROTAC molecule. It encodes biophysical priors to guide the predictions and targets target selection at an early stage.
Bayesian Membership Privacy (BMP) provides a sampling-aware formulation of node-level membership privacy for Graph Neural Networks, incorporating node-dependent priors and treating graph sampling probabilities as part of the privacy analysis. The framework advances privacy guarantees by accounting for graph structure and stochastic training.
EpiFormer uses geometric deep learning to predict antigen–antibody epitopes by jointly modeling antigen and antibody structures. It addresses limitations of chain-wise antibody encoding and severe class imbalance, capturing co-dependent structural features that define binding interfaces.
HYolo integrates hypergraph learning into the YOLO object detector to model high-order contextual relationships among objects and contextual features in IoT settings. Experiments on COCO show significant performance gains.
Identifying and Correcting Label Noise for Robust GNNs via Influence Contradiction introduces ICGNN, which leverages graph structure to detect and correct mislabeled nodes by identifying contradictions in influence, improving robustness to noisy labels.
Fixed Aggregation Features (FAFs) offer a training-free approach that converts graph learning into tabular problems. By computing FAFs and training simple MLPs, this method rivals or even surpasses many state-of-the-art GNNs across 14 benchmarks.
What Structural Inductive Bias Helps Transformers Reason Over Knowledge Graphs? A study with Tabula RASA investigates four independently removable components (sparse adjacency masking, edge-type biases, query scaling, value gating) to isolate their impact. The striking finding is that sparse adjacency masking accounts for most of the improvement, while other components contribute modestly.
SAGE-PTQ (Saliency-Aware Graph-guided Efficient PTQ) proposes an ultra-low-bit post-training quantization framework for LLMs that minimizes hidden scaling costs by separating salient and unsalient weights using distributional statistics and modeling unsalient weights as a sparse graph.
Beyond Vector Similarity analyzes graph-augmented retrieval for industrial knowledge graphs by comparing eight retrieval architectures on a 46-node, 64-edge KG. It shows that five query classes are structurally unreachable by vector-only retrieval, supporting an operator vocabulary thesis that graph computation is necessary for structural queries.
Memory is Reconstructed, Not Retrieved introduces MRAgent, a framework that combines an associative memory graph with active memory reconstruction. Memory is stored as a Cue–Tag–Content graph with semantic bridges to enable dynamic memory access during reasoning rather than static retrieval.
TokenMizer provides a graph-structured session memory system to manage long-horizon LLM contexts. It models the session history as a graph-based memory in a proxy system to preserve structured information beyond the Maximum Effective Context Window.
CausalPOI introduces spatio-temporal graph-based causal modeling for cold-start POI check-in forecasting, aiming to capture functional dependencies and effects of urban interventions beyond simple proximity correlations.
HDST-GNN proposes a heterogeneous dynamic spatiotemporal GNN for multi-object tracking in UAV imagery, with altitude-adaptive edge construction and handling heterogeneous lifecycle states of detections, active tracklets, and lost targets.
Beyond Soft Masks introduces a Hard-Perturbation Mixup Explainer that improves robustness of GNN explanations by extending mixup ideas from soft subgraphs to hard perturbations.
Reducing Hallucinations in Complex Question Answering proposes a simple graph-based Retrieval-Augmented Generation (RAG) approach to reduce hallucinations in complex QA tasks, providing a long-version discussion of the method.
AttackPathGNN reframes Solidity vulnerability detection as reasoning over explicit attack paths using state-interference graphs and conjunction pooling, distinguishing it from prior GNN detectors through architectural choices.
The paper analyzes how entity alignment foundation models underutilize structural context, causing weak cross-KG interaction during encoding and overreliance on coarse similarity in ranking. It proposes ContextEA to inject richer structural context into encoding and apply structure-aware scoring to improve alignment across unseen KG pairs.
This work tackles cold-start item recommendation with an asymmetric graph architecture designed for production constraints. New content receives standalone embeddings for retrieval, and the model also generates device-appropriate embeddings for approximate nearest-neighbor retrieval, bridging semantic and collaborative signals efficiently.
The paper shows how diffusion models can generate graph-like logical rules for knowledge graph reasoning, capturing complex patterns such as cycles and branches beyond simple chains. This approach addresses the combinatorial explosion of rule search by learning generative graph rules that improve reasoning efficiency.
Introducing A2RAG, Adaptive Agentic Graph Retrieval for cost-aware and reliable reasoning, which addresses mixed-difficulty workloads and extraction loss in Graph-RAG. By using adaptive retrieval and agentic evidence routing, it preserves fine-grained qualifiers from source text and improves retrieval efficiency and reliability.
This work provides PAC-Bayesian, margin-based robust generalization analysis for message-passing GNNs under adversarial perturbations, highlighting sensitivity to weight perturbations and offering data-dependent bounds. It extends robustness analysis beyond isotropic Gaussian posteriors and informs robust training strategies.
GraphDETR reframes subgraph detection as a set prediction problem, using a GNN encoder for the target subgraph and a DETR-like, bipartite-matching decoder to predict instances end-to-end, avoiding combinatorial searches inherent in traditional methods.
The paper presents Edge-Aware Curvature Modeling to enhance graph understanding in LLMs by incorporating edge-level information into the graph-text alignment. This edge-centric view improves information propagation across views and strengthens cross-modal reasoning.
WebKnoGraph is an open-source framework for evaluating internal linking strategies on website crawls with GNNs. It models a website as a directed graph and enables offline evaluation of candidate links before deployment to assess effects on authority and semantic coherence.