Flat-Pack Bench - Part Tracking: leaderboard

Flat-Pack Bench (2026) Part Tracking questions (257) on furniture-assembly videos with segmented visual-prompt frames: recover the correspondence of shuffled part IDs between two segmented frames using the video. Multiple-choice accuracy in percent, each model scored in its best of six video and prompt settings; higher is better.

Metric: Track: recover the correspondence of shuffled part IDs between two segmented frames using the video, 257 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 32B (Thinking)45.53#573
2Qwen 2.5 VL 72B Instruct45.14#841
3Perception-LM-8B44.75#932
4InternVL3-78B42.02#759
5Qwen 3 VL 30B A3B Instruct42.02#895
6Qwen 3 VL 235B A22B Instruct39.69#532
7LLaVA-OneVision-72B38.91#1771
8InternVL3-14B37.74#1214
9Qwen 3 VL 4B (Thinking)37.74#1334
10LLaVA-Video-72B35.8#1193
11LLaVA-OneVision-7B34.24#2981
12Qwen 3 VL 8B Instruct33.85#1029
13Qwen 2.5 VL 7B Instruct33.07#1902
14Qwen 2.5 VL 32B Instruct33.07#1062
15Qwen 3 VL 4B Instruct32.68#1332

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

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