AI Paper Atlas

Research map / Graph learning / Graph algorithms and clustering

Neural QAOA$^2$: Differentiable Joint Graph Partitioning and Parameter Initialization for Quantum Combinatorial Optimization

Zubin Zheng, Jiahao Wu, Shengcai Liu

ICML 2026 · Graph learning › Graph algorithms and clustering · approach: Neural combinatorial optimization on graphs

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace heuristic partitioning and random initialization with the learned joint generatorQAOA^2 · In-distribution test (50 instances)mean performance ratio: 0.8869 -> 0.8930 (vs Modularity)Improved
use learned generator on unseen topologiesQAOA^2 · GKA + L (93 OOD instances)average rank: 2.23 -> 1.46 (vs Boundary)Improved
learned generator vs boundary heuristic on Ising spin glassNeural QAOA^2 · L (Ising, OOD)mean performance ratio: 0.8205 -> 0.8160Mixed
learned generator vs modularity on MaxCut WNeural QAOA^2 · W (MaxCut)average rank: 2.23 vs 2.23No clear effect
add test-time adaptation (64 steps)Neural QAOA^2 · in-distribution testmean performance ratio: 0.8892 (vs modularity 0.8869)Improved
add pre-training of the joint generator before test-time adaptation (ablation)Neural QAOA^2 · test-time adaptationmean performance ratioImproved
add Global Topology View to evaluator (ablation)Neural QAOA^2 · 64 TTA stepsmean performance ratio: 0.8819 -> 0.8892Improved
add Partition-Induced Subgraph View to evaluator (ablation)Neural QAOA^2 · 64 TTA stepsmean performance ratio: 0.8797 -> 0.8892Improved
add Quantum Parameter View to evaluator (ablation)Neural QAOA^2 · 64 TTA stepsmean performance ratio: 0.8821 -> 0.8892Improved
add Orthogonal Complement Head instead of unconstrained cluster centers (ablation)Neural QAOA^2 · 64 TTA stepsmean performance ratio: 0.8817 -> 0.8892Improved
add stop-gradient between partition and parameter generation (ablation)Neural QAOA^2 · 64 TTA stepsmean performance ratio: 0.8820 -> 0.8892Improved
use generated topology-aware initial parameters instead of random (ablation)Neural QAOA^2 · 20 QAOA stepsmean performance ratio: 0.8833 -> 0.8892Improved

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