Research map / Architectures and efficiency / Attention and transformers
HATSolver: Learning Gröbner Bases with Hierarchical Attention Transformers
Mohamed Malhou, Ludovic Perret, Kristin E. Lauter
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace flat attention with 2- or 3-level hierarchical attention | base transformer · Gröbner basis n=6, rho=0.33 | sequence accuracy | Improved |
| hierarchical attention for faster training | base transformer · Gröbner basis n=6, rho=0.33 | training steps in 52h: <300K -> ~450K steps | Improved |
| hierarchical attention on larger instances | base transformer · Gröbner basis n=7 rho=0.25 | accuracy: 0% -> 9% (HATSolver-2), 8% (HATSolver-3) | Improved |
| add curriculum learning (ablation) | HATSolver-3 · Gröbner basis n=13, rho=0.9 | exact match accuracy: 33.85 -> 61.2 | Improved |
| HATSolver-3 vs classical algorithms | STD-FGLM/Msolve · n=13 over F7, full density | success rate: 60.8 vs 6.5 (STD-FGLM) and 4.6 (Msolve) | Improved |
| hierarchical attention without curriculum (ablation) | HATSolver-3 · n=13 over F7 | exact match accuracy: 14-34% across test densities | Improved |
| add curriculum learning to base transformer (ablation) | base transformer · Gröbner basis n=7 full density | accuracy: 41% at n=7 rho=1 (prior limit n=5, rho=0.2) | Improved |
| wider (d=1024) vs deeper base transformer (ablation) | base transformer · Gröbner basis n=3 rho=0.5 | training metrics | Improved |
| replace HAT with Hi-Transformer (ablation) | HATSolver-3 · n=13 rho=0.4 | exact match accuracy: HATSolver success vs never learned | Hurt |
| train on non-prime field F16 vs F7 | HATSolver · Gröbner basis n=5 density 20% | exact match accuracy: 52.45 (F7) vs 43.35 (F16) vs 44.01 (F17) | Mixed |
Closest papers by meaning
- Learning to compute Gröbner bases (NeurIPS 2024 · similarity 0.87)
- Computational Algebra with Attention: Transformer Oracles for Border Basis Algorithms (NeurIPS 2025 · similarity 0.86)
- Learning Linear Attention in Polynomial Time (NeurIPS 2025 · similarity 0.84)
- Composing Global Solutions to Reasoning Tasks via Algebraic Objects in Neural Nets (NeurIPS 2025 · similarity 0.83)
- Learning Polynomial Problems with $SL(2, \mathbb{R})$-Equivariance (ICLR 2024 · similarity 0.83)
- Machine Learning meets Algebraic Combinatorics: A Suite of Datasets Capturing Research-level Conjecturing Ability in Pure Mathematics (ICML 2025 · similarity 0.83)
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.
