PinpointQA - Fine-Grained Spatial Description: leaderboard
Fine-grained spatial description on PinpointQA (2026): from 64 frames of an indoor ScanNet++ or ScanNet200 video, the model describes where a small object sits, and a GPT-5.4 judge scores the supporting surface, reference objects, relations, distances and clarity on a 10-point rubric. Mean score scaled to 0-100; higher is better.
Metric: Fine-grained spatial description score (times 100, micro over QA pairs): a GPT-5.4 judge's 10-point rubric (main location and supporting surface 3, reference objects 2, relations 2, distances 2, clarity 1) divided by 10, of the PinpointQA evaluation split (2,019 QA pairs from held-out ScanNet++ and ScanNet200 scenes; 64 uniformly sampled RGB frames per indoor video, no depth or 3D input, greedy decoding); higher is better. Source: arxiv.org. Saturation forecast: Around September 2028. 8 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 32 | #266 |
| 2 | GPT-5.4 | 31 | #135 |
| 3 | Qwen 3 VL 8B Instruct | 28 | #1029 |
| 4 | LLaVA-OneVision-1.5-8B | 26 | #1864 |
| 5 | InternVL3.5-8B-Instruct | 25 | #1891 |
| 6 | Cambrian-S-7B | 24 | #2198 |
| 7 | Spatial-MLLM-820K-4B | 21 | |
| 8 | SenseNova-SI-1.3-InternVL3-8B | 15 | #1610 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=pinpointqa-fine-grained-spatial-description · How It Works · Data refreshed daily, snapshot 2026-11-21.