MM-SAP is a benchmark that systematically evaluates the MLLMs’ self-awareness in perception, which refers to their awareness to recognize what they can and cannot know from the given image.
- [2024.01.15] This repo is made public!
| Rank | Models | BasicVisQA | KnowVisQA | BeyondVisQA | Total |
|---|---|---|---|---|---|
| 1 | GPT-4V | 61.60 ± 2.33 | 83.43 ± 1.40 | 81.96 ± 0.70 | 75.13 ± 1.30 |
| 2 | Qwen-VL-Chat-7b | 70.60 ± 0.68 | 72.11 ± 1.50 | 30.33 ± 0.41 | 57.82 ± 0.63 |
| 3 | InfMLLM-7b | 70.95 ± 1.09 | 55.54 ± 0.50 | 41.03 ± 1.30 | 56.28 ± 0.75 |
| 4 | ShareGPT-4V-7b | 69.70 ± 0.91 | 56.69 ± 1.58 | 41.03 ± 1.36 | 56.19 ± 1.08 |
| 5 | CogVLM-17b | 69.70 ± 0.58 | 66.00 ± 1.26 | 30.49 ± 1.38 | 55.64 ± 0.42 |
| 6 | LLaVA-13b | 68.25 ± 1.56 | 58.11 ± 1.43 | 34.02 ± 0.90 | 53.81 ± 0.39 |
| 7 | ShareGPT-4V-13b | 68.00 ± 1.74 | 60.29 ± 1.11 | 30.43 ± 0.54 | 53.22 ± 0.60 |
| 8 | LLaVA-7b | 62.00 ± 0.85 | 54.23 ± 1.92 | 30.27 ± 0.85 | 49.12 ± 0.60 |
@article{wang2024mm,
title={MM-SAP: A Comprehensive Benchmark for Assessing Self-Awareness of Multimodal Large Language Models in Perception},
author={Wang, Yuhao and Liao, Yusheng and Liu, Heyang and Liu, Hongcheng and Wang, Yu and Wang, Yanfeng},
journal={arXiv preprint arXiv:2401.07529},
year={2024}
}