MirrorBench - Task Success Rate: 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. Task success rate, the share of episodes in which the hand reaches the mark, over all levels and bodies, from 0 to 1; higher is better.

Metric: Task success rate (0-1), the share of episodes in which the hand reaches the mark, 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 February 2027. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.50.17#261
2Gemini 2.5 Pro0.155#276
3GPT-4.10.115#505
4GLM-4.5V0.096#754
5Gemini 2.5 Flash0.064#463
6GPT-4o0.061#751
7Qwen-VL-Max (MirrorBench checkpoint unspecified)0.053
8Qwen 2.5 VL 72B0.023#781
9GPT-4o Mini0.008#1667
10LLaVA-Next-LLaMA3-8B0.004#5654
11InternVL3.5-4B0.004#1930
12Qwen 2.5 VL 32B0.004#941
13Qwen 2.5 VL 7B0#1862
14LLaVA-v1.5-7B0#7630
15Qwen 2.5 VL 3B0#2639

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

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