BARISTA - Phrase Grounding: leaderboard

BARISTA (2026) is an egocentric benchmark of 185 coffee-preparation videos with verified per-frame scene graphs, from which zero-shot language tasks for vision-language models are derived. This board is COCO-style box mAP in percent for localizing the region a phrase refers to; higher is better.

Metric: COCO-style mAP (%, IoU 0.5 to 0.95, times 100) of the boxes a model returns for a referring phrase, over 1,000 frozen single frames on BARISTA, an egocentric coffee-preparation video dataset with verified per-frame scene graphs, zero-shot with shared task prompts, temperature 0.7 (GPT-5.4 mini at its fixed sampling), low thinking effort for the closed models, mean of three runs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 FlashLow53.4#175
2Qwen 3.5 27B51.8#375
3Gemma 4 31B43.2#382
4Qwen 3.5 9B35#828
5GPT-5.4 MiniLow6.8#406

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

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