AI Paper Atlas

Research map / Graph learning / GNN architectures

Neutral-atom quantum features as complementary structural encodings for graph learning

Igor Sokolov, Mathieu Garrigues, Matthias Hecker, Lorenzo Moro, Shaheen Acheche

NeurIPS 2026 · Graph learning › GNN architectures · approach: Physics-inspired GNNs · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes using measurement features from neutral-atom quantum processors running graph-dependent Rydberg Hamiltonians as complementary structural encodings for GNNs.

Problem

Classical structural encodings for GNNs (random walks, Laplacian spectra) capture only specific structural statistics of a graph, leaving room for alternative graph descriptors.

Approach

Neutral-atom QPUs realize graph-dependent Rydberg Hamiltonians, so quantum measurement-derived features probing graph topology are used as structural encodings complementary to classical ones for graph learning.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add AFM quantum features to RRWPRRWP encoding, DeepSet probe · ZINC intrinsic probe: degreeMAE (x1e-2): 14.4 -> 1.18Improved
add AFM quantum features to RRWPRRWP encoding, DeepSet probe · ZINC intrinsic probe: node min cycleMAE (x1e-2): 0.73 -> 0.36Improved
add FM to RRWP+AFMRRWP+AFM · ZINC intrinsic probe: eccentricityMAE: 1.77 -> 1.47 (eccentricity); best on local tasks otherwiseMixed
add quantum features to RRWP in GNN backbones (champion hybrid vs RRWP)six GNN backbones · MUTAG and ZINC subsetF1 / MAE: ties or improves in 10 of 12 cells; sign test p = 0.019Improved
add AFM_stat+FM_stat to RRWPGAT · ZINC subsetMAE: 0.223 -> 0.209Improved
add quantum features to RRWP on GeneralConvGeneralConv · ZINC subsetMAE: 0.145 vs 0.145 (RRWP+Laplacian)No clear effect
use quantum features alone instead of RRWP (ablation)GNN backbones · MUTAG and ZINC subsetF1 / MAE: ZINC GT MAE 0.113 -> 0.163 (AFM)Hurt
replace emulated with QPU hardware features (AFM_stat hybrid)RRWP + AFM_stat · MUTAGmacro F1: within 0.01 F1 on all six backbonesNo clear effect
replace emulated with hardware AFM features (hybrid)RRWP + AFM · MUTAGmacro F1: GCN 0.878 -> 0.837; within 0.014 on five of sixMixed
use gated fusion instead of concatenation (ablation)ESA with RWSE*+quantum features · MUTAGbinary F1: RWSE*+AFM+FM 0.903 -> 0.922Improved
concatenate quantum features to RWSE* in ESAESA · MUTAGbinary F1: 0.905 -> 0.899 (AFM_stat)No clear effect
use AFM+FM quantum features instead of classical features on ladder graphslogistic regression probe · Synthetic ladder graphsaccuracy: 0.641 -> 0.875Improved

Experimental setup

Baselines: random walk encodings Laplacian spectral encodings

Builds on: random walk positional encoding Laplacian positional encoding

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy