Research map / Label-efficient and robust learning / Recognition and long-tail
The Natural Geometry of Code: Hyperbolic Representation Learning for Program Reasoning
Weilin Zhou
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 |
|---|---|---|---|
| use hyperbolic (Lorentz) operations instead of Euclidean counterparts (ablation) | HypeCodeNet · BigCloneBench | F1-score: 0.923 -> 0.940 | Improved |
| add geodesic-based (hyperbolic) attention (ablation) | HypeCodeNet · BigCloneBench | F1-score: 0.932 -> 0.940 | Improved |
| add curvature annealing (ablation) | HypeCodeNet · BigCloneBench | F1-score: 0.935 -> 0.940 | Improved |
| increase embedding dimension from 32 to 768 (ablation) | HypeCodeNet · BigCloneBench | F1-score: 0.928 -> 0.940 | Improved |
| increase embedding dimension from 64 to 128 (ablation) | HypeCodeNet · BigCloneBench | F1-score: 0.934 -> 0.938 | Improved |
| increase embedding dimension from 128 to 256 (ablation) | HypeCodeNet · BigCloneBench | F1-score: 0.938 -> 0.939 | No clear effect |
| hyperbolic GNN instead of best Euclidean hybrid baseline | HypeCodeNet vs CodeFORMER · BigCloneBench | F1-score: 0.928 -> 0.940 | Improved |
| hyperbolic GNN instead of best Euclidean hybrid baseline | HypeCodeNet vs CodeFORMER · POJ-104 | accuracy: 0.974 -> 0.981 | Improved |
| hyperbolic GNN instead of best Euclidean hybrid baseline | HypeCodeNet vs CodeFORMER · CodeXGLUE completion (Python) | accuracy (%): 44.0 -> 45.0 | Improved |
| hyperbolic GNN instead of best Euclidean hybrid baseline | HypeCodeNet vs CodeFORMER · CodeXGLUE completion (Java) | accuracy (%): 41.4 -> 42.2 | Improved |
| hyperbolic GNN instead of best Euclidean hybrid baseline | HypeCodeNet vs CodeFORMER · GitHub Java call-graph link prediction | AUC: 0.915 -> 0.965 | Improved |
| hyperbolic GNN instead of best Euclidean hybrid baseline | HypeCodeNet vs CodeFORMER · GitHub Java call-graph link prediction | Hits@10: 0.768 -> 0.820 | Improved |
Closest papers by meaning
- HELM: Hyperbolic Large Language Models via Mixture-of-Curvature Experts (NeurIPS 2025 · similarity 0.87)
- Low-distortion and GPU-compatible Tree Embeddings in Hyperbolic Space (ICML 2025 · similarity 0.86)
- Hyperbolic Embeddings of Supervised Models (NeurIPS 2024 · similarity 0.85)
- Hyperbolic Fine-Tuning for Large Language Models (NeurIPS 2025 · similarity 0.85)
- Learning Along the Arrow of Time: Hyperbolic Geometry for Backward-Compatible Representation Learning (ICML 2025 · similarity 0.84)
- Exploiting Code Symmetries for Learning Program Semantics (ICML 2024 · similarity 0.84)
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.
