← Home

Daily arXiv Papers

Graph Neural Networks · Graph Learning · LLM × Graph

Showing 10 papers for 2026-02-26

MINAR: Mechanistic Interpretability for Neural Algorithmic Reasoning
GNN

MINAR introduces Mechanistic Interpretability for Neural Algorithmic Reasoning, a framework to identify circuit-like components inside GNNs that implement classical algorithms. It aims to provide efficient circuit discovery and mechanistic explanations for neural algorithmic reasoning. This helps understand how algorithmic reasoning is carried out in neural models.

RABot: Reinforcement-Guided Graph Augmentation for Imbalanced and Noisy Social Bot Detection
GNN Graph Learning

RABot proposes Reinforcement-guided Graph Augmentation to improve social bot detection under severe class imbalance and noisy deceptive links. It uses multi-granularity graph augmentations guided by reinforcement learning to produce informative views that boost GNN-based detectors.

A Causal Graph-Enhanced Gaussian Process Regression for Modeling Engine-out NOx
Graph Learning

We propose a causal graph-enhanced Gaussian Process Regression for modeling engine-out NOx, providing a probabilistic framework that captures uncertainty and encodes domain-relevant causal structure. This enables more reliable real-time monitoring and diagnostics beyond traditional deterministic models.

Training speedups via batching for geometric learning: an analysis of static and dynamic algorithms
GNN Graph Learning

This work analyzes batching algorithms for geometric learning, investigating how static and dynamic batching affect training time and model performance in graph neural networks. The results offer practical insights into when batching speeds training without sacrificing accuracy.

Temporal Knowledge-Graph Memory in a Partially Observable Environment
Knowledge Graph

We introduce Room Environment v3, a configurable partially observable setting whose hidden state is a RDF knowledge graph and whose observations are RDF triples. The agent can extend observations into a temporal KG to store and reason about evolving state. The benchmark supports configurable world dynamics and memory tasks.

Nonstabilizerness Estimation using Graph Neural Networks
GNN

We present a Graph Neural Network approach to estimate nonstabilizerness in quantum circuits, measured by stabilizer Renyi entropy (SRE). We formulate three supervised learning tasks, from classification to regression, and show that GNNs can effectively estimate nonstabilizerness, enabling scalable assessment of quantum resources.

EExApp: GNN-Based Reinforcement Learning for Radio Unit Energy Optimization in 5G O-RAN
GNN Graph Learning

We propose EExAPP, a GNN-based reinforcement learning xApp for 5G Open Radio Access Network that jointly optimizes radio unit sleep scheduling and distributed unit resource slicing. The system uses a dual-actor-dual-critic PPO architecture to learn policies. Experiments show energy savings while maintaining quality of service.

PiPNN: Ultra-Scalable Graph-Based Nearest Neighbor Indexing
Graph Theory

We introduce PiPNN, an ultra-scalable graph-based nearest neighbor indexing method. It avoids the costly search-driven construction bottleneck of beam-search based methods by using HashPrune online pruning, enabling fast construction while preserving high query quality.

I/O Optimizations for Graph-Based Disk-Resident Approximate Nearest Neighbor Search: A Design Space Exploration
Graph Theory

This paper studies I/O optimizations for disk-resident ANN, presenting an I/O-first framework that jointly considers memory layout, disk layout, and search algorithm. A page-level complexity model explains how page locality and path length affect reads, and experiments on four public datasets validate the approach and quantify the effects of different design choices.

Towards Autonomous Graph Data Analytics with Analytics-Augmented Generation
Graph Learning

We argue for Analytics-Augmented Generation AAG, a paradigm that treats analytical computation as a first-class component in end-to-end graph data analytics with LLM agents. The approach goes beyond retrieval or code generation, focusing on intent-to-execution translation, deliberate graph construction, and reliable execution across diverse graph algorithms.