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..
Geometry-Aware Collaborative Multi-Solutions Optimizer for Model Fine-Tuning with Parameter Efficiency
Authors: Van-Anh Nguyen, Trung Le, Mehrtash Harandi, Ehsan Abbasnejad, Thanh-Toan Do, Dinh Phung
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across transfer learning, few-shot learning, and domain generalization show that our proposed approach consistently outperforms existing Bayesian methods, delivering strong performance with affordable computational overhead and offering a practical solution by updating only a small subset of parameters. The code for our method is at https://github.com/anh-ntv/GAC-MSO |
| Researcher Affiliation | Academia | Van-Anh Nguyen Department of Data Science and AI Monash University, Australia EMAIL |
| Pseudocode | Yes | The training algorithm is presented in Supplementary. |
| Open Source Code | Yes | The code for our method is at https://github.com/anh-ntv/GAC-MSO |
| Open Datasets | Yes | VTAB-1k dataset [53] consists of 19 distinct datasets, which are grouped into three categories: Natural, Specialized, and Structured. Each dataset contains only 1,000 images for training, making the task challenging due to the limited amount of data. Additionally, the images show significant variation in data distribution across the datasets, further complicating the learning process. |
| Dataset Splits | Yes | Each dataset contains only 1,000 images for training, making the task challenging due to the limited amount of data. [...] In this section, we extend our analysis to a few-shot learning setting by varying the number of training samples (shots) per class across 1, 2, 4, 8, and 16. [...] the model is fine-tuned on a subset of the Image Net-1K dataset [13], which includes 16 samples per class. |
| Hardware Specification | No | We do not include discussion of the resource. |
| Software Dependencies | No | The paper mentions various methods and models but does not provide specific software library names with version numbers used for implementation. |
| Experiment Setup | Yes | Detail of the experimental setting is presented in Appendix A, which includes the backbone, how to set up multiple particles, the kernel function, and trade-off parameters. [...] We conduct experiments using four particles for our GAC-MSO and all baselines, except full finetuning (FFT), Adam W, and SAM, for which we use a single particle consistent with standard Lo RA-based fine-tuning of foundation models. Each particle is randomly initialized at the start. |