AI Paper Atlas

Research map / Graph learning / GNN architectures

Can Computational Reducibility Lead to Transferable Models for Graph Combinatorial Optimization?

Semih Cantürk, Thomas Sabourin, Frederik Wenkel, Michael Perlmutter, Guy Wolf

ICML 2026 · Graph learning › GNN architectures · approach: Expressive equivariant message passing · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

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 local message passing (GCN, GIN, GATv2) with the GCON wavelet filter bank (ablation)energy-based unsupervised solver · RB-small MVCMVC size: 221.56 -> 211.69Improved
use Hamiltonian loss rather than quadratic first-term loss for MIS (ablation)GCON · MIS on RB-smallindependent set size: 15.5 -> 18.04Improved
fine-tune whole GNN with inverted output head for MIS to MVC transfer (ablation)GCON · RB-small MIS->MVCMVC size: 211.69 -> 211.56Improved
freeze GNN backbone and only retrain head for MIS<->MVC (ablation)GCON · RB-small MIS and MVCMIS size / MVC size: MIS 18.12 -> 17.68; MVC 211.69 -> 212.46Hurt
invert output head instead of resetting when backbone frozen (ablation)GCON · RB-small MIS<->MVCMIS size: 17.68 -> 17.69No clear effect
fine-tune full model (instead of frozen) when transferring MIS to MaxClique (ablation)GCON · RB-small MaxClique, G featuresclique size: 16.12 -> 16.55Improved
add 3-layer multi-head attention on frozen MIS backbone for MaxClique (ablation)GCON · RB-small MaxCliqueclique size: 16.12 -> 16.13No clear effect
fine-tune on complement graph with MIS-pretrained backbone (true reduction) (ablation)GCON · RB-small MaxCliqueclique size: 15.52 (frozen) -> 16.82 (fine-tuned)Improved
pretrain on the other five tasks, then fine-tune for 20 epochs on K-coloringGCON · BA-small leave-one-outcoloring violations: 61.92 -> 24.19Improved
pretrain on other tasks then fine-tune MDSGCON · BA-small leave-one-outMDS size: 35.57 -> 36.15No clear effect
pretrain on MDS, MIS, Coloring then fine-tune MaxClique, MaxCut, MVC for 20 epochsGCON · BA-smallMVC size: 141.43 -> 139.70Improved
pretrain on MDS, MIS, Coloring then fine-tune MaxCliqueGCON · BA-smallclique size: 4.31 -> 4.32 (full: 4.43)No clear effect

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