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..
Plug-and-play Feature Causality Decomposition for Multimodal Representation Learning
Authors: Ye Liu, Zihan Ji, Hongmin Cai
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on 9 existing multimodal methods and 5 multimodal datasets prove the effectiveness of FCD. |
| Researcher Affiliation | Academia | Ye Liu Zihan Ji Hongmin Cai School of Future Technology South China University of Technology Guangzhou, China 511442 EMAIL EMAIL |
| Pseudocode | Yes | The pseudo code is in Appendix Section B. Algorithm 1: Enforce Full Rank in Py Torch style ... Algorithm 2: Training process of our process method in Py Torch style ... Algorithm 3: Measure-preserving Linear Layer Forward |
| Open Source Code | No | We will make the code public after acceptance. |
| Open Datasets | Yes | We design and conduct our experiments on 5 widely used multimodal datasets, i.e., CMUMOSI [41], CMU-MOSEI [42], MSVA-Single [27], UPMC-Food101 [3], and HFM [4]. |
| Dataset Splits | Yes | Table 5 summarizes the training, validation and test subset splits following [15; 43]. Table 5: Datasets splits (train, validation (val), and test) in our experiments. DATASET TRAIN VAL TEST OVERALL CMU-MOSI 1284 229 686 2199 CMU-MOSEI 16326 1871 4659 22856 MVSA-SINGLE 1555 518 519 2592 UPMC FOOD101 62971 5000 22715 90686 HFM 19816 2410 2409 24635 |
| Hardware Specification | Yes | Our experiments are conducted on 4 NVIDIA RTX 4090 24GB GPUs with Py Torch [29] framework. |
| Software Dependencies | No | Our experiments are conducted on 4 NVIDIA RTX 4090 24GB GPUs with Py Torch [29] framework. |
| Experiment Setup | Yes | The main hyper-parameters in this paper consist of two parts, i.e., the hyper-parameters in each multi-modal intermediate fusion method (e.g., learning rate, batch size) and the ones of FCD (λ1, λ2, λ3). The hyper-parameters of FCD and their sensitive analysis are summarized in Appendix Section E, and please refer to the original papers for other hyper-parameters of each base model. Moreover, we report the computational overhead and complexity analysis in Appendix Section F. We fix the temperature parameter τ = 0.07 in Info NCE loss for the most common cases. ... Table 6: Hyper-parameters (λ1, λ2 and λ3) settings in our experiments. |