1: \begin{abstract}

2: 

3: Optimizing the performance of many objectives (instantiated by tasks or clients) jointly with a few Pareto stationary solutions (models) is critical in machine learning. However, previous multi-objective optimization

4: methods often focus on a few number of objectives and cannot scale to many objectives that outnumber the solutions, 

5: leading to either subpar performance or ignored objectives.  

6: We introduce ``{\bf M}any-{\bf o}bjective multi-{\bf s}olution {\bf T}ransport (\name)'', a framework that finds multiple diverse solutions in the Pareto front of many objectives.  

7: Our insight is to seek multiple solutions, each performing as a domain expert and focusing on a specific subset of objectives while collectively covering all of them. 

8: \name formulates the problem as a bi-level optimization of weighted objectives for each solution, where the weights are defined by an optimal transport between the objectives and solutions. 

9: Our algorithm ensures convergence to Pareto stationary solutions for complementary subsets of objectives.   

10: On a range of applications in federated learning, multi-task learning, and mixture-of-prompt learning for LLMs, 

11: \name distinctly outperforms strong baselines, delivering high-quality,

12: diverse solutions that profile the entire Pareto frontier, thus

13: ensuring balanced trade-offs across many objectives.

14: \end{abstract}

15: