CAPO: Confidence Aware Preference Optimization Learning for Multilingual Preferences

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Pokharel, Rhitabrat, Tao, Yufei, Agrawal, Ameeta
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909897626484736
author Pokharel, Rhitabrat
Tao, Yufei
Agrawal, Ameeta
author_facet Pokharel, Rhitabrat
Tao, Yufei
Agrawal, Ameeta
contents Preference optimization is a critical post-training technique used to align large language models (LLMs) with human preferences, typically by fine-tuning on ranked response pairs. While methods like Direct Preference Optimization (DPO) have proven effective in English, they often fail to generalize robustly to multilingual settings. We propose a simple yet effective alternative, Confidence-Aware Preference Optimization (CAPO), which replaces DPO's fixed treatment of preference pairs with a dynamic loss scaling mechanism based on a relative reward. By modulating the learning signal according to the confidence in each preference pair, CAPO enhances robustness to noisy or low-margin comparisons, typically encountered in multilingual text. Empirically, CAPO outperforms existing preference optimization baselines by at least 16% in reward accuracy, and improves alignment by widening the gap between preferred and dispreferred responses across languages.
format Preprint
id arxiv_https___arxiv_org_abs_2511_07691
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CAPO: Confidence Aware Preference Optimization Learning for Multilingual Preferences
Pokharel, Rhitabrat
Tao, Yufei
Agrawal, Ameeta
Computation and Language
Artificial Intelligence
Preference optimization is a critical post-training technique used to align large language models (LLMs) with human preferences, typically by fine-tuning on ranked response pairs. While methods like Direct Preference Optimization (DPO) have proven effective in English, they often fail to generalize robustly to multilingual settings. We propose a simple yet effective alternative, Confidence-Aware Preference Optimization (CAPO), which replaces DPO's fixed treatment of preference pairs with a dynamic loss scaling mechanism based on a relative reward. By modulating the learning signal according to the confidence in each preference pair, CAPO enhances robustness to noisy or low-margin comparisons, typically encountered in multilingual text. Empirically, CAPO outperforms existing preference optimization baselines by at least 16% in reward accuracy, and improves alignment by widening the gap between preferred and dispreferred responses across languages.
title CAPO: Confidence Aware Preference Optimization Learning for Multilingual Preferences
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2511.07691