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..
Efficiently Maintaining the Multilingual Capacity of MCLIP in Downstream Cross-Modal Retrieval Tasks
Authors: Fengmao Lyu, Jitong Lei, Guosheng Lin, Desheng ZHENG, Jianyang Zhang, Tianrui Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental evaluations across diverse datasets validate the effectiveness and scalability of our framework, demonstrating robust multilingual retention across languages. Our codes are available at https://github.com/tiggers23/Ta PCL-Ci PCL. |
| Researcher Affiliation | Collaboration | Fengmao Lv1, Jitong Lei1, Guosheng Lin2, Desheng Zheng3,4, Jianyang Zhang5 , Tianrui Li1 1Southwest Jiaotong University, 2Nanyang Technological University 3Southwest Petroleum University, 4Kash Institute of Electronics and Information Industry, 5University of Electronic Science and Technology of China EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods and formulas, e.g., in sections 4.1, 4.2, and 4.3, but does not present a structured pseudocode or algorithm block. |
| Open Source Code | Yes | Our codes are available at https://github.com/tiggers23/Ta PCL-Ci PCL. |
| Open Datasets | Yes | We evaluate our proposed methods on MSCOCO36 [10] and XM3600 [31], both of which are cross-modal retrieval benchmarks involving 36 parallel corpora. |
| Dataset Splits | Yes | For MSCOCO36 and XM3600, we respectively randomly select 5000 and 2880 samples for training and set English as the source language. |
| Hardware Specification | Yes | The experiments are conducted on Linux with NVIDIA 3090 GPUs |
| Software Dependencies | Yes | We adopt the Open CLIP Vi T-B-32-XLM-Roberta-Base model [5] as our pre-trained multi-lingual CLIP model. For Ci PCL, we use M2M100-1.2B [26] to translate, and we translate the critical top-1 word (we also report the detailed results with the top-2 and top-3 words in A.4.) into target languages. |
| Experiment Setup | Yes | The experiments are conducted on Linux with NVIDIA 3090 GPUs, using Adam W optimizer with cosine learning rate scheduler, and the initial learning rate is 1e 4. We set the batch size to 128 and α = 0.2, β = 0.8. The model is trained for 10 epochs with PCL-base on MSCOCO36, and for 20 epochs on XM3600. |