Interactive Visual Grounding (Full): leaderboard
The Interactive Visual Grounding benchmark (2026) casts a vision-language model as the Matcher in a reference game: a simulated Director describes target items drawn from human-grounded image sets of dogs, baskets and tangrams, and the model must pick each target from a grid, asking clarification questions where the protocol allows. In the Full protocol the Matcher gets the full description with optional clarification questions. Score is the share of targets matched correctly, in percent; higher is better.
Metric: Matcher accuracy (%; share of target items the model selects correctly given the full description with optional clarification questions; 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 February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 89.9 | #94 |
| 2 | GPT-5.5 (2026-04-23) | 69.5 | #33 |
| 3 | GPT-5.4 (2026-03-05) | 69.4 | #87 |
| 4 | Gemini 3.1 Flash Lite (Preview) | 65.3 | #418 |
| 5 | Gemma 4 31B (IT) | 60 | #342 |
| 6 | Qwen 3 VL 32B Instruct | 50.5 | #585 |
| 7 | GPT-5.4 Mini (2026-03-17) | 49 | #287 |
| 8 | Qwen 3 VL 8B Instruct | 35 | #981 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=interactive-visual-grounding-full · How It Works · Data refreshed daily, snapshot 2026-11-13.