To Trust Or Not To Trust Your Vision-Language Model's Prediction

Hao Dong¹  Moru Liu²  Jian Liang^3,4  Eleni Chatzi¹  Olga Fink⁵

¹ETH Zurich, ²Technical University of Munich, ³NLPR & MAIS, IInstitute of Automation, Chinese Academy of Sciences, ⁴University of Chinese Academy of Sciences, ⁵EPFL

• arXiv 2025 •

---

Illustration of TrustVLM's mechanism. Initially, the incorrect prediction receives a higher confidence score than the correct one, indicating overconfidence. By performing verification in the image embedding space, this overconfidence is mitigated. As a result, the final confidence score is significantly higher for the correct prediction than for the incorrect one.

Prerequisites

Environment

The code was tested using Python 3.10.13, torch 2.3.1+cu121 and NVIDIA GeForce RTX 3090. More dependencies are in requirement.txt.

Datasets

We suggest downloading all datasets to a root directory, and renaming the directory of each dataset as suggested in ${ID_to_DIRNAME} in ./data/datautils.py. Please refer to this page to download and prepare datasets accordingly.

Run TrustVLM-D

bash ./test_trustvlm.sh

Run TrustVLM*-D

bash ./test_trustvlm_v2.sh

Citation

If you find our work useful in your research please consider citing our paper:

@article{dong2025trustvlm,
    title={To Trust Or Not To Trust Your Vision-Language Model's Prediction},
    author={Dong, Hao and Liu, Moru and Liang, Jian and Chatzi, Eleni and Fink, Olga},
    journal={arXiv preprint arXiv:2505.23745},
    year={2025}
}

Acknowledgements

We thank the authors of TPT and CoOp for their open-source implementation and instructions on data preparation.