ESI-Bench - Mirror Spatial Relations: leaderboard
Embodied spatial intelligence (ESI-Bench, Hong et al., 2026): questions about spatial relations revealed by a mirror and the real view together in BEHAVIOR-1K scenes in the OmniGibson simulator, designed so that a fixed view is not enough. The vision-language model explores actively, moving the camera and manipulating objects, before answering; accuracy in percent, higher is better.
Metric: Accuracy (%; Spatial Relations questions in the Specular Reflection category, active multi-view exploration). Source: arxiv.org. Saturation forecast: Around February 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 54.8 | #163 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=esi-bench-mirror-spatial-relations · How It Works · Data refreshed daily, snapshot 2026-11-12.