Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Advancing Interpretability of CLIP Representations with Concept Surrogate Model
Authors: Nhat Hoang-Xuan, Xiyuan Wei, Wanli Xing, Tianbao Yang, My T. Thai
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive quantitative experiments across multiple datasets validate the faithfulness of EXPLAIN-R, demonstrating that the surrogate representation accurately preserves the predictive behavior of the original CLIP model. We establish via a user study that EXPLAIN-R produces explanations considered relevant to the input image, sufficiently complete to explain the model s capabilities, and useful for overall model comprehension. |
| Researcher Affiliation | Academia | Nhat Hoang-Xuan University of Florida Xiyuan Wei Texas A&M University Wanli Xing University of Florida Tianbao Yang Texas A&M University My T. Thai University of Florida Corresponding author. Email: EMAIL |
| Pseudocode | Yes | We present the pseudocode and full derivation in Appendix B.2. Algorithm 1: Algorithm for Surrogate Learning |
| Open Source Code | Yes | We provide code along with instructions to reproduce our results in the supplemental materials. |
| Open Datasets | Yes | We use the COCO 2017 [43] validation set, Flickr30k [44], SUN397 [45] test set, and Image Net validation set [46] to study the CLIP model s behavior. |
| Dataset Splits | Yes | For the Flickr30k dataset, we explain on the full dataset. For the SUN397 dataset, we use the first official testing split. For the COCO 2017 and Image Net dataset, we use the validation split. |
| Hardware Specification | Yes | All experiments are performed on a single A100 GPU. |
| Software Dependencies | No | We use the Amsgrad variant of the Adam W optimizer with learning rate 10 3 and weight decay 10 6. The paper does not explicitly mention any specific software versions (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We train the surrogate with Algorithm 1 for 150 epochs, with batch size 1024, default temperature τ = 0.1, and γ1 = γ2 = 0.9. The optimizer used is Adam W with learning rate 10 3. |