Interactive Visual Grounding (Static): leaderboard

Metric: Matcher accuracy (%; share of target items the model selects correctly given the full human-produced description, no interaction; micro-averaged over 536 target items of four human-grounded image sets (dogs, two basket sets, tangrams), with GPT-5.4 as the simulated Director answering from the human description; a 10-turn budget per target, utterances capped at 50 words). Source: arxiv.org. Saturation forecast: Around December 2026. 8 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)87.5
2Gemini 3.1 Flash Lite (Preview)67.8
3GPT-5.4 (2026-03-05)66.1
4GPT-5.565.8
5Gemma 4 31B (IT)56
6Qwen 3 VL 32B Instruct48.8
7GPT-5.4 Mini47.9
8Qwen 3 VL 8B Instruct33.6

Interactive version: theaggregate.ai/benchmark?slug=interactive-visual-grounding-static · How It Works · Data refreshed daily, snapshot 2026-09-29.