PointQ-Bench - Quality Reasoning: leaderboard

Open-ended point-cloud quality reports on 3,083 point clouds from real scans, simulated distortions and AI-generated assets (PointQ-Bench, 2026), with 2D models seeing six rendered views and 3D models sampled points, scored 0-10 as the sum of five 0-2 judge ratings (structure, specificity, faithfulness, coherence, quality accuracy) against an expert reference. Higher is better.

Metric: SSFRQ-5D score (0-10): sum of five 0-2 judge ratings (structural sufficiency, specificity, faithfulness, reasoning coherence, quality accuracy) of the open-ended quality report against an expert reference, zero-shot over 3,083 point clouds (authentic scans, simulated distortions, AI-generated assets), temperature 0; 2D models see six rendered views, native 3D models 8,192 sampled points; free-form answers mapped to the label space by an LLM parser; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5Low8.41#178
2GPT-5 MiniLow8.21#364
3Qwen 3.5 9BNon-reasoning7.89#828
4Gemma 3 12B7.77#2059
5Qwen 3.5 27BNon-reasoning7.77#375
6Claude Haiku 4.57.76#583
7InternVL3.5-14B7.76#1504
8Mistral Small 3.27.73#1641
9InternVL3.5-8B7.72#1604
10ShapeLLM-13B6.32
11ShapeLLM-7B5.93
12PointLLM-13B4.49
13PointLLM-7B4.31
14MiniGPT-3D3

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

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