Flat-Pack Bench - Temporal Localization: leaderboard

Flat-Pack Bench (2026) Temporal Localization questions (103) on furniture-assembly videos with segmented visual-prompt frames: identify the events just before or after the state shown in the visual prompt. Multiple-choice accuracy in percent, each model scored in its best of six video and prompt settings; higher is better.

Metric: TLoc: identify the events just before or after the state shown in the visual prompt, 103 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: Around February 2027. 30 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-553.4#178
2Qwen 3 VL 32B Instruct46.6#606
3Gemini 2.5 Pro44.66#276
4Gemini 3.1 Pro (Preview)43.69#99
5Gemini 2.5 Flash41.75#463
6InternVL3-78B39.81#759
7InternVL3-38B37.86#939
8Video-LLaVA-7B35.92#3879
9Qwen 3 VL 4B Instruct33.01#1332
10Qwen 3 VL 8B (Thinking)33.01#910
11Perception-LM-3B32.04
12Qwen 3 VL 8B Instruct30.1#1029
13Qwen 2.5 VL 72B Instruct30.1#841
14LLaVA-OneVision-7B30.1#2981
15Flat-Pack Bench VideoRefer (checkpoint unspecified)30.1

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-temporal-localization · How It Works · Data refreshed daily, snapshot 2026-11-21.