Perception Test (MCQA)
Multimodal · 2023-05-23
The Perception Test from Google DeepMind probes perception and reasoning over purpose-filmed real-world videos that target memory, abstraction, physics and semantics skills. This row is its multiple-choice video question-answering task on the test split, scored by top-1 accuracy over three answer options; higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 2.5 Pro | 78.4 |
| Gemini 2.5 Flash | 75.1 |
| Qwen2.5-VL-72B-Instruct | 73.2 |
| Qwen2.5-VL-7B-Instruct | 70.5 |
| Gemini 1.5 Pro 002 | 69.4 |
| Gemini 2.0 Flash | 68.8 |
| Qwen2 VL 72B Instruct | 68.0 |
| Gemini 2.0 Flash-Lite | 67.5 |
| Qwen2.5-VL-3B-Instruct | 66.9 |
| Gemini 1.5 Flash 002 | 66.5 |
| GPT-4.1 | 64.8 |
| Qwen2 VL 2B Instruct | 53.9 |