AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

HONet: Data-Efficient Learning for Exact Cover Tasks via Hypergraph Optimization

Pengyang Huang, Zirui Zhuang, Haifeng Sun, Qi Qi, Jingyu Wang, Jianxin Liao

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Differentiable optimization layers

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
add relaxation gap loss Lgap (GCL) (ablation)HONet · Sudoku 9x9 (20-22 Givens)whole-board accuracy (%): 99.70 -> 100.00Improved
add relaxation gap loss Lgap (GCL) (ablation)HONet · Sudoku 9x9 (26-28 Givens)relaxation gap: ~10-2 -> ~10-5Improved
add relaxation gap loss Lgap (GCL) (ablation)HONet · Sudoku 9x9 (26-28 Givens)inference time: 56.51ms -> 11.45msImproved
use hypergraph encoder instead of Graph Transformer (ablation)HONet · Sudoku 9x9 (26-28 Givens)epochs to peak validation: 18 -> 4Improved
use hypergraph encoder instead of Graph Transformer (ablation)HONet · Sudoku 9x9 (20-22 Givens)whole-board accuracy (%): 99.50 -> 100.00Improved
add neural network objective shaping instead of direct projection of clues (ablation)HONet · Sudoku 9x9 (20-22 Givens)whole-board accuracy (%): 30.20 -> 100.00Improved
add auxiliary warm-start loss Lwarm and reduce lambda (ablation)HONet · Sudoku 9x9 (17-19 Givens)whole-board accuracy (%): ~3 -> 91.80Improved
train on 20-22 Givens only, test zero-shot on 17-19 GivensHONet · Sudoku 9x9 (OOD)whole-board accuracy (%): 87.50Improved
reduce training set to 500 samples (~5%) (ablation)HONet · Sudoku 9x9 (Medium/Hard)whole-board accuracy (%): > 99No clear effect
reduce training set to 100 samples (~1%) (ablation)HONet · Sudoku 9x9 (Medium/Hard)whole-board accuracy (%): ~90Hurt
replace SATNet solver layer with HONetVisual Sudoku model · Visual Sudokutest accuracy (%): 63.20 -> 98.40Improved
replace SATNet with HONetVisual Sudoku model · Visual Sudokuepochs to peak: 99 -> 24Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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