analiza aiqbench
This paper introduces TPO, a novel framework addressing distributional shifts in Direct Preference Optimization (DPO) by modeling hierarchical uncertainties across group, prompt, and preference levels. By integrating these multi-level constraints into a single tractable objective, the method offers improved robustness for safe large language model deployment compared to existing approaches that handle uncertainty sources individually.
TPO: Tri-level Distributionally Robust Learning for OOD Direct Preference Optimization
arXiv:2610.28229v1 Announce Type: new Abstract: Direct Preference Optimization (DPO) minimizes empirical risk over observed preference comparisons, implicitly assuming that the training preference distribution matches the deployment preference distribution. In practice, this i.i.d. assumption often fails, as preference distribution shift exhibits a natural hierarchical structure across three levels: group-level shift, prompt-level shift, and preference-level shift. Such multi-level distributional mismatch poses a critical challenge for safe LLM deployment, yet existing robust DPO methods typically either model uncertainty over the joint preference distribution or focus on individual sources of shift, leaving the hierarchical structure of OOD preference shifts underexplored. We propose TPO, a tri-level distributionally robust framework for OOD preference alignment, which jointly models all three levels of distributional uncertainty through a hierarchically structured rectangular ambiguity set. By instantiating it with Wasserstein distance, we derive a tractable and interpretable objective that combines a closed-form preference-level margin penalty, prompt-level adversarial embedding perturbation, and worst-case group reweighting, each governed by an independent radius. We establish finite-sample convergence guarantees and empirically validate the effectiveness of TPO over existing robust DPO methods.
Abstract:Direct Preference Optimization (DPO) minimizes empirical risk over observed preference comparisons, implicitly assuming that the training preference distribution matches the deployment preference distribution. In practice, this i.i.d. assumption often fails, as preference distribution shift exhibits a natural hierarchical structure across three levels: group-level shift, prompt-level shift, and preference-level shift. Such multi-level distributional mismatch poses a critical challenge for safe LLM deployment, yet existing robust DPO methods typically either model uncertainty over the joint preference distribution or focus on individual sources of shift, leaving the hierarchical structure of OOD preference shifts underexplored. We propose TPO, a tri-level distributionally robust framework for OOD preference alignment, which jointly models all three levels of distributional uncertainty through a hierarchically structured rectangular ambiguity set. By instantiating it with Wasserstein distance, we derive a tractable and interpretable objective that combines a closed-form preference-level margin penalty, prompt-level adversarial embedding perturbation, and worst-case group reweighting, each governed by an independent radius. We establish finite-sample convergence guarantees and empirically validate the effectiveness of TPO over existing robust DPO methods.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2610.28229 [cs.LG]
(or arXiv:2610.28229v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.28229 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 267 cuvinte
Subiecte conexe
Leveraging Discrete Function Decomposability for Scientific Design
arXiv · 9 sept. 2026
First-order Constrained Trilevel Optimization Over Distributed Networks for Robust Coreset Selection
arXiv · 31 iul. 2026
Constrained Diffusion Models with Primal-Dual Inference
arXiv · 14 oct. 2026
Using Non-Lipschitz Signum-based Functions for Distributed Optimization and Machine Learning: Trade-off Between Con-vergence Rate and Optimality Gap
arXiv · 4 aug. 2026