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..
Enhancing Compositional Reasoning in CLIP via Reconstruction and Alignment of Text Descriptions
Authors: Jihoon Kwon, Kyle Min, Jy-yong Sohn
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Sec. 4, we provide experiments demonstrating the READ method is effective across a wide range of compositional reasoning benchmarks. Specifically, we introduce READ-CLIP, a VLM derived by applying the READ method to the pre-trained CLIP model [45], which achieves the state-of-the-art performance on five compositional reasoning benchmarks. |
| Researcher Affiliation | Collaboration | Jihoon Kwon Seoul National University EMAIL Kyle Min Oracle EMAIL Jy-yong Sohn Yonsei University EMAIL |
| Pseudocode | No | The paper describes methods and objectives in prose and uses diagrams (Figure 1, Figure 2) to illustrate the architecture and loss components, but it does not contain explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Codes are available at this Git Hub repository. |
| Open Datasets | Yes | We use the MS-COCO dataset [35] for all experiments. We evaluate READ-CLIP and the baselines on five benchmarks Whats Up [24], CREPE [37], VALSE [40], Sugar Crepe [19], and Sugar Crepe++ [11] each designed to assess a different aspect of compositional reasoning. |
| Dataset Splits | Yes | We follow training practices established in prior work on compositional reasoning [38, 54, 64, 68], using a 100K subsample with the Karpathy split [25], 5 training epochs, a batch size of 256, and the Vi T-B/32 architecture. |
| Hardware Specification | Yes | All experiments are conducted using a single A100 40GB GPU. |
| Software Dependencies | No | The paper mentions fine-tuning models using the Huggingface transformers [61] library, but does not provide specific version numbers for this or any other key software components, such as Python or PyTorch, that would be necessary for full reproducibility. |
| Experiment Setup | Yes | We follow training practices established in prior work on compositional reasoning [38, 54, 64, 68], using a 100K subsample with the Karpathy split [25], 5 training epochs, a batch size of 256, and the Vi T-B/32 architecture. The Adam W optimizer is used with a learning rate of 1.0 10 5, cosine annealing schedule, 50 warmup steps, and a weight decay of 0.1. Finally, the weighting factors in Eq. 8 are set to α = 0.1 and β = 0.5. |