Many-Objective Multi-Solution Transport
Authors: Ziyue Li, Tian Li, Virginia Smith, Jeff Bilmes, Tianyi Zhou
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On a range of applications in federated learning, multi-task learning, and mixture-of-prompt learning for LLMs, Mos T distinctly outperforms strong baselines, delivering high-quality, diverse solutions that profile the entire Pareto frontier, thus ensuring balanced trade-offs across many objectives. ... In all applications (Section 6), Mos T finds diverse high-quality solutions on the Pareto front, consistently outperforming various strong baselines in terms of average accuracy and other popular metrics on the quality of multiple solutions, without extra computation cost. ... Section 6 is titled 'MOST APPLICATIONS' and contains subsections like 'EXPERIMENTAL SETUP', 'FEDERATED LEARNING', 'MULTI-TASK LEARNING', 'MIXTURE-OF-PROMPT LEARNING', and 'ABLATION STUDIES AND COMPARISON WITH OTHER BASELINES', all detailing empirical evaluations and comparisons. |
| Researcher Affiliation | Academia | 1University of Maryland 2University of Chicago 3Carnegie Mellon University 4University of Washington, Seattle EMAIL, EMAIL, EMAIL, EMAIL. All listed affiliations are universities, and email domains are .edu. |
| Pseudocode | Yes | Algorithm 1 Many-Objective Multi-Solution Transport |
| Open Source Code | Yes | Project: https://github.com/tianyi-lab/Mos T |
| Open Datasets | Yes | We conduct experiments on synthetic data and Federated Extended MNIST (FEMNIST) (Cohen et al., 2017; Caldas et al., 2018) ... Office-Caltech10 (Saenko et al., 2010; Griffin et al., 2007) and Domain Net (Peng et al., 2019) ... three datasets from the Super GLUE benchmark (Wang et al., 2019) ... toy ZDT problem set (Zitzler et al., 2000) ... German credit dataset (Asuncion & Newman, 2007). |
| Dataset Splits | Yes | For the generated dataset, we conduct our experiments using a train-validation-test split ratio of 6:2:2. ... We randomly sample 128 instances from the training dataset and evenly partition the original validation dataset to form both the validation and test datasets. |
| Hardware Specification | Yes | Table 9: Runtime (sec) comparisons for all methods on federated learning datasets, performed on a single Nvidia RTX A5000 platform. ... Table 12: Runtime (sec) comparisons for all methods on ZDT datasets, performed on a single Nvidia RTX A4000 platform. |
| Software Dependencies | No | The paper mentions software components like 'IPOT (Xie et al., 2020)', 'Frank-Wolfe algorithms (Fujishige, 1980)', 'T5-base model', but does not provide specific version numbers for these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | The learning rates are swept from {0.005, 0.01, 0.05, 0.1} without decaying throughout the training process. ... We run 400 epochs in total. ... We run 400 epochs for training. Learning rates are swept from {0.08, 0.1}. ... conducting 20 epochs of training and sweeping learning rates from {0.08, 0.1}. |