Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 21 papers for 2026-08-20
This paper investigates turning abductive explanations into global logical rules for node classification in SGCs (Simplified Graph Convolutional Networks). It notes that explanatory subgraphs tied to individual nodes can be redundant and limit rule generality, and it proposes a logic-based framework to extract broader, node-agnostic rules.
This work presents a general framework for infusing semantic knowledge into spatio-temporal traffic forecasting. By constructing semantic subgraphs around traffic sensors from a knowledge graph (e.g., Wikidata), the method augments sensor data with contextual knowledge to improve predictions.
HyPE-GT introduces a framework that learns hyperbolic positional encodings for Graph Transformers to better capture hierarchical graph structure. By contrast with traditional Euclidean encodings, the approach encodes nodes in hyperbolic space to reflect nested relationships and improve attention.
This work tackles oversquashing by adopting asynchronous message passing across layers. At each layer, node features are updated asynchronously, guided by centrality metrics to preserve long-range information transmission, resulting in a model-agnostic improvement without rewiring or extra parameters.
This paper examines the privacy risks of community inference by Graph Neural Networks and explores defense mechanisms to conceal sensitive group-level structure. It studies a defensive setting where a network or defender seeks to obfuscate or suppress detectable community signals to protect privacy.
GEqTrain is a configuration-driven framework for retargeting equivariant GNNs across 3D scientific tasks. It separates dataset semantics, model composition, and training objectives, mapping raw data to typed fields and assembling stacks, losses, and workflows via Hydra configurations.
NICE proposes Scale-Stable Perturbations for GNN explanations, addressing distribution shift caused by perturbations. By analyzing the perturbation mechanism and critiquing Element-wise Masking, NICE aims to produce more reliable explanations without undermining predictions.
EquiPocket is an E(3)-equivariant geometric GNN for ligand binding site prediction. It enforces 3D geometric symmetries to robustly model protein structures, addressing rotation sensitivity and irregular surfaces, and improving binding-site prediction.
This survey reframes self-evolving agents as dynamic graph transformers, arguing that agent states, memories, tools, and execution structures form evolving graphs. It advocates treating graphs as evolving substrates rather than static supports for agent functions.
RDFdL integrates RDF with Differential Dynamic Logic to enable reasoning over static knowledge and continuous dynamics. It provides a syntaxic representation of differential equations and associated reasoning rules, bridging symbolic and dynamic reasoning for cyber-physical systems.
GenEx constructs codon co-occurrence graphs from gene sequences to capture contextual interdependencies not reflected in linear sequences. This graph-based representation enables new features for SARS-CoV-2 variant detection beyond traditional alignment methods.
FairGlucose provides a fairness benchmark for CGM forecasting with a balanced, multi-strata cohort and extensive samples. Evaluating 33 models reveals subgroup disparities that population-level validation masks, highlighting the need for subgroup-aware evaluation in CGM forecasting.
SIDScope offers a diagnostic resource for Semantic-ID interfaces in generative recommender systems. It normalizes item-tokenizer mappings, verifies provenance, profiles structural properties, and tracks how mapping revisions affect generated outputs.
rEDMRec distills reasoning from large language models into an editable Experience Memory for recommendations, enabling reuse, inspection, and adaptation across requests as user tastes drift.
Pre-Compiled Pipeline Shards enable distributed LLM inference on Intel AI PC fleets by splitting models layer-wise into pre-compiled OpenVINO graphs. This pipeline-parallel setup reduces memory bottlenecks and accelerates inference across multiple machines.
A framework and prototype for a navigable map of datasets in Engineering Design and Systems Engineering addresses data fragmentation by cataloging datasets across the design lifecycle, supporting findability, reproducibility, and method validation.
From Sequence to Structure proposes Relational Uncertainty Propagation for LLM agents, enabling uncertainty to propagate along relations over sequences. This enhances robustness and failure detection beyond token-level signals.
Learning Random Geometric Graphs Drawn in Probabilistic Metric Spaces develops a data-driven method to construct RGGs in a probabilistic metric space, using a distribution over a disparity variable to define connectivity. It is applicable across data types and sizes.
A unifying relational perspective on expressive lottery tickets generalizes the Strong Expressive Lottery Ticket hypothesis to multi-relational and temporal graphs, showing sparse subnetworks can retain Weisfeiler–Lehman expressivity in RGNNs/TGNNs.
Visual-Aware Representation of Web Pages presents a platform using the FitLayout rendering tool to capture visual and layout information for ML on web content, enabling visual-aware representations.
SINDI introduces an efficient index for sparse vector approximate MIPS, designed for production-grade retrieval in Retrieval-Augmented Generation (RAG) systems. It analyzes bottlenecks of existing inverted- and graph-based approaches—redundant distance computations, frequent random memory accesses, and the difficulty of SIMD acceleration with compressed sparse storage—and proposes a sparse inverted non-redundant indexing scheme to reduce redundancy and memory traffic. This leads to faster MIPS throughput in realistic workloads.