FAM-Bench - Comparative Ranking (CoT): leaderboard

Comparative dish ranking in FAM-Bench (2026), a food-as-medicine benchmark: given a health condition and four candidate dishes with recipe text and images, the model ranks them by suitability, under the CoT prompting mode, in which the model first writes short reasoning steps. The score is top-1 accuracy in percent, chance 25; higher is better.

Metric: Top-1 accuracy (%) on Task 2: rank four candidate dishes for a health condition and match the gold most suitable dish, CoT prompting mode (the model first writes short reasoning steps), 1,000 comparative queries over four candidate dishes with recipe text and image, 13 health conditions, nutrition-expert verified, greedy decoding with JSON output; chance 25; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.635.7#173
2Gemini 2.5 Pro34.6#276
3GPT-5.433.1#135
4Gemma 3 12B (IT)31.3#1887
5Qwen 3 VL 8B Instruct29.3#1029

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

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