MirrorBench: leaderboard

MirrorBench (2026) adapts the mirror self-recognition test to multimodal models: in a simulated scene the model moves a hand to touch a mark on its own body that it can only see in a mirror. Overall score over four prompt levels and human and robot bodies. The score averages task success, step-wise improvement and final and peak completion ratios (at most 1; negative when the hand ends farther away than it started); higher is better.

Metric: Overall average score (mean of task success rate, step-wise improvement ratio, final completion ratio and peak completion ratio, each at most 1; the final completion ratio is negative when the hand ends farther from the mark than it started), over the four prompt levels and both body settings, on MirrorBench's Isaac Sim mirror self-recognition scenes (7 human and robot bodies, 6 hands, 6 marks; the agent moves its hand with six directional actions to touch a mark on its own body seen in a mirror; step budget the shortest path plus 10), temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro0.442#276
2Gemini 2.5 Flash0.359#463
3GPT-4.10.27#505
4GPT-4o0.25#751
5Qwen-VL-Max (MirrorBench checkpoint unspecified)0.238
6Claude Sonnet 4.50.222#261
7Qwen 2.5 VL 72B0.154#781
8GLM-4.5V0.098#754
9Qwen 2.5 VL 32B0.089#941
10InternVL3.5-4B0.02#1930
11GPT-4o Mini-0.14#1667
12LLaVA-Next-LLaMA3-8B-0.21#5654
13LLaVA-1.6-7B-0.277#4955
14InternVL3.5-1B-0.302#6112
15Qwen 2.5 VL 7B-0.306#1862

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mirrorbench · How It Works · Data refreshed daily, snapshot 2026-11-21.