Research map / 3D vision / Face and shape
Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2
Yuri Chervonyi, Trieu H. Trinh, Miroslav Olšák, Xiaomeng Yang, Hoang H. Nguyen, Marcelo Menegali, Junehyuk Jung, Junsu Kim, Vikas Verma, Quoc V. Le, Thang Luong
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 |
|---|---|---|---|
| upgrade AG1 to AG2 (new language, engine, LM, SKEST search) | AlphaGeometry · IMO-AG-50 | solve rate: 54% -> 84% | Improved |
| add multiple search trees with knowledge sharing (SKEST) (ablation) | AG2 single search tree · IMO-AG-50 | problems solved: 38 -> 42 | Improved |
| add multiple search trees with knowledge sharing (SKEST) (ablation) | AG2 single search tree · IMO-AG-30 | problems solved: 28 -> 30 | No clear effect |
| replace DDAR1 by DDAR2 (faster algorithm, C++) | symbolic engine · 25 IMO problems | runtime (s): 1179.57 -> 3.44711 | Improved |
| replace DDAR1 by DDAR2 symbolic engine | DDAR · IMO-AG-50 | problems solved: 14 -> 16 | Improved |
| add language model to symbolic engine (ablation) | AG2 · IMO-AG-50 | problems solved: 16 -> 42 | Improved |
| sample with temperature 1.0 and 32 samples instead of greedy decoding (ablation) | AG2 LM · 26 IMO problems needing auxiliary constructions | problems solved: 2 -> 9 | Improved |
| use a custom small-vocabulary tokenizer vs large Gemini tokenizer (ablation) | AG LM · IMO 2000-2024 | solve rate | No clear effect |
| train on natural language instead of AG domain language (ablation) | AG LM · IMO 2000-2024 | solve rate | No clear effect |
| fine-tune math-pretrained 3.3B Gemini instead of training from scratch (ablation) | AG LM · IMO-AG-50 | problems solved | No clear effect |
| add image input of the diagram (multimodal) (ablation) | AG2 LM alone · IMO problems | solve rate | No clear effect |
| increase beam size or samples beyond beam 128, depth 4, 32 samples (ablation) | AG2 single search tree · IMO problems | problems solved | No clear effect |
Closest papers by meaning
- PuzzleFusion++: Auto-agglomerative 3D Fracture Assembly by Denoise and Verify (ICLR 2025 · similarity 0.76)
- Omni-MATH: A Universal Olympiad Level Mathematic Benchmark for Large Language Models (ICLR 2025 · similarity 0.76)
- Proving Olympiad Algebraic Inequalities without Human Demonstrations (NeurIPS 2024 · similarity 0.76)
- SBSC: Step-by-Step Coding for Improving Mathematical Olympiad Performance (ICLR 2025 · similarity 0.76)
- SolidGeo: Measuring Multimodal Spatial Math Reasoning in Solid Geometry (NeurIPS 2025 · similarity 0.76)
- Fully Dynamic Euclidean Bi-Chromatic Matching in Sublinear Update Time (ICML 2025 · similarity 0.76)
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.
