GeoCodeBench: leaderboard
Metric: Unit-test pass rate (%) on all 100 problems, GeoCodeBench's 100 fill-in-the-function problems from 47 recent CVPR, ICCV and ICLR 3D-vision repositories; the model gets the structured paper text, the source file with the target function masked and an execution template, one greedy generation; score is the share of the problem's 10 tool-generated, expert-reviewed unit tests that pass, averaged over problems; higher is better. Source: arxiv.org. Saturation forecast: Around July 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 36.6 | #91 |
| 2 | Claude Sonnet 4.5 | 31.1 | #138 |
| 3 | Gemini 2.5 Pro | 30.4 | #145 |
| 4 | Kimi K2 | 30.4 | #236 |
| 5 | Seed-1.6 | 26.9 | #257 |
| 6 | Qwen 3 Coder 480B A35B Instruct | 23.5 | #302 |
| 7 | DeepSeek R1 | 21 | #245 |
| 8 | Llama 3.1 405B Instruct | 14.3 | #447 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=geocodebench · How It Works · Data refreshed daily, snapshot 2026-10-11.