Point-Bench
Multimodal · 2025-05-15
Point-Bench, from PointArena, tests language-guided pointing with 982 image-instruction pairs in five categories: spatial, affordance, counting, steerability and reasoning. A response succeeds when its point lands inside the target mask (for counting, when it places one point on each target and no extras), and the score is the success rate.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-VL-2B | 73.8 |
| Gemma 4 12B IT | 73.4 |
| GPT-5.6 Luna | 66.8 |
| Qwen3.5-4B | 56.7 |
| Qwen3.5-2B | 54.0 |
| Gemini 3.5 Flash-Lite | 51.4 |
| Gemma 4 E2B IT | 14.3 |