Enregistré dans:
Détails bibliographiques
Auteurs principaux: Sharifnassab, Arsalan, Salehkaleybar, Saber, Ghiassian, Sina, Kanoria, Surya, Schuurmans, Dale
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:https://arxiv.org/abs/2405.00747
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866909335176609792
author Sharifnassab, Arsalan
Salehkaleybar, Saber
Ghiassian, Sina
Kanoria, Surya
Schuurmans, Dale
author_facet Sharifnassab, Arsalan
Salehkaleybar, Saber
Ghiassian, Sina
Kanoria, Surya
Schuurmans, Dale
contents We propose Soft Preference Optimization (SPO), a method for aligning generative models, such as Large Language Models (LLMs), with human preferences, without the need for a reward model. SPO optimizes model outputs directly over a preference dataset through a natural loss function that integrates preference loss with a regularization term across the model's entire output distribution rather than limiting it to the preference dataset. Although SPO does not require the assumption of an existing underlying reward model, we demonstrate that, under the Bradley-Terry (BT) model assumption, it converges to a softmax of scaled rewards, with the distribution's "softness" adjustable via the softmax exponent, an algorithm parameter. We showcase SPO's methodology, its theoretical foundation, and its comparative advantages in simplicity, computational efficiency, and alignment precision.
format Preprint
id arxiv_https___arxiv_org_abs_2405_00747
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Soft Preference Optimization: Aligning Language Models to Expert Distributions
Sharifnassab, Arsalan
Salehkaleybar, Saber
Ghiassian, Sina
Kanoria, Surya
Schuurmans, Dale
Machine Learning
Artificial Intelligence
We propose Soft Preference Optimization (SPO), a method for aligning generative models, such as Large Language Models (LLMs), with human preferences, without the need for a reward model. SPO optimizes model outputs directly over a preference dataset through a natural loss function that integrates preference loss with a regularization term across the model's entire output distribution rather than limiting it to the preference dataset. Although SPO does not require the assumption of an existing underlying reward model, we demonstrate that, under the Bradley-Terry (BT) model assumption, it converges to a softmax of scaled rewards, with the distribution's "softness" adjustable via the softmax exponent, an algorithm parameter. We showcase SPO's methodology, its theoretical foundation, and its comparative advantages in simplicity, computational efficiency, and alignment precision.
title Soft Preference Optimization: Aligning Language Models to Expert Distributions
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2405.00747