IGG - Reveal Performed: leaderboard
IGG paper v1: goal-only interactive GUI grounding, 601 tasks in 16 controllable apps. Screenshot and action history only, no DOM/accessibility/target ID; 1,440x900 viewport, deterministic element-ID click success, reset before each task, eight actions, one run and greedy decoding. Exact open checkpoints are Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen3-VL-8B-Instruct and mPLUG/GUI-Owl-1.5-8B-Instruct. GPT-5.5 API effort not separately disclosed. Oracle intervention performs the reveal before model grounding. GPT-5.4 Mini uses low effort. Metric: Grounding accuracy (%).
Metric: Grounding accuracy (%). Source: arxiv.org. Saturation forecast: Around January 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini (Low) | 90.4 | #215 (GPT-5.4 Mini) |
| 2 | Qwen 3 VL 8B Instruct | 81.9 | #421 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=igg-reveal-performed · How It Works · Data refreshed daily, snapshot 2026-10-27.