Flat-Pack Bench - Temporal Localization: leaderboard
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 July 2027. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 53.4 |
| 2 | Qwen 3 VL 32B Instruct | 46.6 |
| 3 | Gemini 2.5 Pro | 44.66 |
| 4 | Gemini 3.1 Pro (Preview) | 43.69 |
| 5 | Gemini 2.5 Flash | 41.75 |
| 6 | InternVL3-78B | 39.81 |
| 7 | InternVL3-38B | 37.86 |
| 8 | Qwen 3 VL 4B Instruct | 33.01 |
| 9 | Qwen 3 VL 8B (Thinking) | 33.01 |
| 10 | Qwen 3 VL 8B Instruct | 30.1 |
| 11 | Qwen 2.5 VL 72B Instruct | 30.1 |
| 12 | Qwen 2.5 VL 32B Instruct | 29.13 |
| 13 | Qwen 3 VL 235B A22B Instruct | 25.24 |
| 14 | Qwen 3 VL 4B (Thinking) | 25.24 |
| 15 | Qwen 3 VL 30B A3B Instruct | 22.33 |
Interactive version: theaggregate.ai/benchmark?slug=flat-pack-bench-temporal-localization · How It Works · Data refreshed daily, snapshot 2026-10-07.