MirrorBench - Level 3 (Self-Referential Mirror Recognition): 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. Level 3, where neither the mirror nor the fact that the mark is on the body of the agent itself is mentioned. The score averages task success, step-wise improvement and final and peak completion ratios (at most 1; the final completion ratio is negative when the hand ends farther away than it started); higher is better.

Metric: 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) at Level 3, where neither the mirror nor the fact that the mark is on the body of the agent itself is mentioned, over 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 Pro0.342#280
2Gemini 2.5 Flash0.235#466
3Qwen-VL-Max (MirrorBench checkpoint unspecified)0.222
4GPT-4.10.108#506
5InternVL3.5-4B0.082#1946
6Qwen 2.5 VL 72B0.006#782
7Qwen 2.5 VL 32B-0.051#944
8GPT-4o-0.058#754
9Claude Sonnet 4.5-0.154#267
10LLaVA-1.6-7B-0.248#4990
11GLM-4.5V-0.288#756
12Qwen 2.5 VL 7B-0.307#1878
13Qwen 2.5 VL 3B-0.348#2664
14GPT-4o Mini-0.392#1680
15InternVL3.5-1B-0.398#6140

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mirrorbench-level-3-self-referential-mirror-recognition · How It Works · Data refreshed daily, snapshot 2026-11-22.