Flat-Pack Bench - Part Mating: leaderboard

Flat-Pack Bench (2026) Part Mating questions (87) on furniture-assembly videos with segmented visual-prompt frames: decide whether two parts are connected in the final assembly. Multiple-choice accuracy in percent, each model scored in its best of six video and prompt settings; higher is better.

Metric: Mate: decide whether two parts are connected in the final assembly, 87 questions; multiple-choice accuracy (%, regex exact match) on furniture-assembly videos from IMaW with one or two segmented visual-prompt frames, zero-shot with greedy decoding (GPT-5 at its default); each row is the model's best of up to six settings (key-frame or trimmed video, mixed-media, collage or concat visual prompt) chosen by overall micro-average accuracy, the paper's main-table protocol; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 30 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 4B (Thinking)59.77#1354
2LLaVA-OneVision-72B57.47#1782
3Qwen 3 VL 4B Instruct56.32#1348
4LLaVA-Video-72B56.32#1211
5Qwen 2.5 VL 32B Instruct54.02#1065
6InternVL3-38B52.87#948
7LLaVA-Video-7B52.87#2303
8Video-LLaVA-7B51.72#3904
9Flat-Pack Bench VideoRefer (checkpoint unspecified)51.72
10Gemini 3.1 Pro (Preview)49.43#102
11GPT-549.43#185
12Qwen 3 VL 30B A3B Instruct49.43#903
13InternVL3-14B48.28#1227
14Qwen 3 VL 32B (Thinking)47.13#577
15Qwen 3 VL 8B (Thinking)44.83#924

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=flat-pack-bench-part-mating · How It Works · Data refreshed daily, snapshot 2026-11-22.