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..
Preference Optimization on Pareto Sets: On a Theory of Multi-Objective Optimization
Authors: Abhishek Roy, Geelon So, Yian Ma
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | Abstract and introduction clearly states our main result with explicit rate. As this is a theoretical paper, this summarizes our main contribution. ... In this work, we provide a principled and efficient way to select a decision vector from the Pareto set of a set of objectives f1, . . . , fn given an additional preference function f0. The primary motivation is to seek the most preferred solution from a large model like LLM that is pretrained to satisfy a number of desiderata. A main contribution of this work is to provide a geometrically-meaningful notion of (approximate) preference stationarity. This is non-trivial due to the non-smoothness and non-convexity of the Pareto set. We achieve this by reformulating the constraint set as the Pareto manifold instead of the Pareto set. We also provide a simple algorithm that achieves ε0-approximate stationarity with iteration complexity of O(ε 2 0 ), under both first-order and dueling feedback. |
| Researcher Affiliation | Academia | Abhishek Roy Texas A&M University EMAIL Geelon So UC San Diego EMAIL Yi-An Ma UC San Diego EMAIL |
| Pseudocode | Yes | Algorithm 1 Pareto majorization-minimization (PMM) ... Algorithm 2 Approximation of preference gradient with dueling feedback |
| Open Source Code | Yes | The code is contained in the supplementary materials and at https://github. com/geelon/preference-pareto. |
| Open Datasets | No | This is a theoretical paper focused on algorithm design and convergence analysis, not empirical validation against specific, publicly available datasets. The NeurIPS checklist confirms 'NO experiment.' for dataset-related questions. |
| Dataset Splits | No | This is a theoretical paper focused on algorithm design and convergence analysis. It does not conduct experiments on datasets requiring splits. The NeurIPS checklist confirms 'NO experiment.' |
| Hardware Specification | No | This is a theoretical paper. While it includes visualizations, these are not extensive empirical experiments for which specific hardware would typically be detailed. The NeurIPS checklist states 'No experiment.' |
| Software Dependencies | No | This is a theoretical paper. While code is provided, specific software dependencies with version numbers are not detailed within the main text or appendices for replication. The NeurIPS checklist states 'No experiment.' |
| Experiment Setup | Yes | I Details to Visualization of Toy Examples: Hyperparameters used using the PNG algorithm by Ye and Liu (2022) include the Pareto-set approximation threshold e = 0.001, learning rate ξ = 0.01, and regularization parameter αt = 0.01. These were not particularly tuned. The PMM algorithm visualize in this example alternates between updating xk and βk. Each update step for βk is implemented by an approximate gradient descent, making use of ˆ x , followed by a projection back onto the simplex. The learning rate for β chosen here was 0.1. Then, the update step for xk consists of a single step of gradient descent on fβk with learning rate of 0.1. A total of 1500 iterations was run. These hyperparameters were not specifically tuned. |