Interactive Visual Grounding (Partial): leaderboard

Metric: Matcher accuracy (%; share of target items the model selects correctly given only the least informative attribute of the description, so the Matcher must ask for the rest; 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)89
2GPT-5.571.4
3GPT-5.4 (2026-03-05)65.5
4Gemma 4 31B (IT)58.4
5Gemini 3.1 Flash Lite (Preview)54.5
6Qwen 3 VL 32B Instruct44
7GPT-5.4 Mini37.1
8Qwen 3 VL 8B Instruct35.3

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