FoodSense: leaderboard

FoodSense (2026) asks vision-language models to predict, from a food photo alone, how it would taste, smell, feel and sound, as 1-5 ratings compared with the mean of about 22 human raters on 438 test images. Lin's concordance correlation coefficient averaged over the four senses (-1 to 1, shown times 100); higher is better.

Metric: Average over taste, smell, texture and sound of Lin's concordance correlation coefficient between the model's predicted 1-5 rating and the mean human rating (about 22 annotators per image) on FoodSense's 438 held-out test food images, zero-shot with the authors' sensory-assessment prompt, ratings parsed by regex; range -1 to 1, shown times 100; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemma 3 27B (IT)13.6#1324
2Qwen 2.5 VL 32B Instruct12.4#1062
3LLaVA-v1.6-34B11.3#5950
4InternVL2.5-26B7.8#2156
5food-Llama-3.2-11B-Vision-Instruct5.5

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

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