Interactive Visual Grounding (None): leaderboard

Metric: Matcher accuracy (%; share of target items the model selects correctly given no description, the Director answering only yes, no or unclear; 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
1GPT-5.559.7
2Gemini 3.1 Pro (Preview)56.7
3GPT-5.4 (2026-03-05)43.2
4Gemini 3.1 Flash Lite (Preview)37.9
5Gemma 4 31B (IT)26.4
6Qwen 3 VL 32B Instruct23.9
7Qwen 3 VL 8B Instruct21.6
8GPT-5.4 Mini13.5

Interactive version: theaggregate.ai/benchmark?slug=interactive-visual-grounding-none · How It Works · Data refreshed daily, snapshot 2026-09-29.