StateSight - Occluded Cube Counting: leaderboard
StateSight (2026) is a procedurally generated benchmark for latent spatial-state reconstruction in vision-language models, with deterministic oracle labels and exact-match scoring over 300 single-image prompts per task. Accuracy on counting all unit cubes in an isometric tower including hidden supports, in percent, higher is better.
Metric: Accuracy (%; counting all unit cubes in an isometric tower including hidden supports; 300 procedurally generated single-image prompts with deterministic oracle labels, exact-match, answer-only prompt). Source: arxiv.org. Saturation forecast: Around November 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Non-reasoning) | 33.3 | #48 (GPT-5.5) |
| 2 | Claude Sonnet 5 (Non-reasoning) | 18.7 | #108 (Claude Sonnet 5) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=statesight-occluded-cube-counting · How It Works · Data refreshed daily, snapshot 2026-11-12.