PinpointQA: leaderboard

Small object-centric spatial understanding in indoor videos on PinpointQA (2026), built from ScanNet++ and ScanNet200 scenes: from 64 frames the model works through four progressively harder tasks, from verifying that a small target is present to predicting its location as structured JSON. Mean of the four task scores, scaled to 0-100; higher is better.

Metric: Average score (0-1 task scores, times 100): arithmetic mean of the four task micro scores (target presence verification, nearest reference identification, fine-grained spatial description, structured spatial prediction) 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 November 2028. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Kimi K2.542#266
2Qwen 3 VL 8B Instruct39#1029
3GPT-5.438#135
4LLaVA-OneVision-1.5-8B35#1864
5InternVL3.5-8B-Instruct34#1891
6Cambrian-S-7B34#2198
7SenseNova-SI-1.3-InternVL3-8B32#1610
8Spatial-MLLM-820K-4B26

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=pinpointqa · How It Works · Data refreshed daily, snapshot 2026-11-21.