NutriBench-Kitchen: leaderboard
NutriBench-Kitchen (Wei et al., 2026) tests whether vision-language models can track ingredient states through long cooking videos and combine them with recipe and nutrition knowledge, over 1,500 manually verified questions on 160 videos. The score is the unweighted mean in percent of ten cells, five task families (ingredient entry, memory management, recipe query, long-term and short-term planning) each in an easy and a hard regime, with accuracy on multiple-choice and mean relative accuracy on numerical answers; higher is better.
Metric: Average score (%; unweighted mean of the ten task-regime cells: five task families (ingredient entry, memory management, recipe query, long-term and short-term planning) in an easy and a hard regime; NutriBench-Kitchen: 1,500 manually verified questions on 160 cooking videos (145 HD-EPIC segments and 15 self-recorded videos with measured ingredient quantities); accuracy on multiple-choice questions and mean relative accuracy on numerical ones; zero-shot with the same chain-of-thought prompt for every model). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 61.1 | #262 |
| 2 | Qwen 2.5 VL 7B Instruct | 50.8 | #1842 |
| 3 | GPT-4o | 47.7 | #741 |
| 4 | GLM-4.1V-9B (Thinking) | 29.4 | #1061 (GLM-4.1V-9B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=nutribench-kitchen · How It Works · Data refreshed daily, snapshot 2026-11-12.