Offline and Online KL-Regularized RLHF under Differential Privacy

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wu, Yulian, Thareja, Rushil, Vepakomma, Praneeth, Orabona, Francesco
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911211835097088
author Wu, Yulian
Thareja, Rushil
Vepakomma, Praneeth
Orabona, Francesco
author_facet Wu, Yulian
Thareja, Rushil
Vepakomma, Praneeth
Orabona, Francesco
contents In this paper, we study the offline and online settings of reinforcement learning from human feedback (RLHF) with KL-regularization -- a widely used objective function in large language model alignment -- under the $ε$ local differential privacy ($ε$-LDP) model on the label of the human preference. In the offline setting, we design an algorithm based on the principle of pessimism and derive a new suboptimality gap of $\tilde{O}(1/[(e^ε-1)^2 n])$ on the KL-regularized objective under single-policy concentrability. We also prove its optimality by providing a matching lower bound where $n$ is the sample size. In the online setting, we are the first one to theoretically investigate the problem of KL-regularized RLHF with LDP. We design an optimism-based algorithm and derive a logarithmic regret bound of $O(d_{\mathcal{F}}\log (N_{\mathcal{F}}\cdot T) /(e^ε-1)^2 )$, where $T$ is the total time step, $N_{\mathcal{F}}$ is cardinality of the reward function space $\mathcal{F}$ and $d_{\mathcal{F}}$ is a variant of eluder dimension for RLHF. As a by-product of our analysis, our results also imply the first analysis for online KL-regularized RLHF without privacy. We implement our algorithm in the offline setting to verify our theoretical results and release our open source code at: https://github.com/rushil-thareja/PPKL-RLHF-Official.
format Preprint
id arxiv_https___arxiv_org_abs_2510_13512
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Offline and Online KL-Regularized RLHF under Differential Privacy
Wu, Yulian
Thareja, Rushil
Vepakomma, Praneeth
Orabona, Francesco
Machine Learning
Artificial Intelligence
In this paper, we study the offline and online settings of reinforcement learning from human feedback (RLHF) with KL-regularization -- a widely used objective function in large language model alignment -- under the $ε$ local differential privacy ($ε$-LDP) model on the label of the human preference. In the offline setting, we design an algorithm based on the principle of pessimism and derive a new suboptimality gap of $\tilde{O}(1/[(e^ε-1)^2 n])$ on the KL-regularized objective under single-policy concentrability. We also prove its optimality by providing a matching lower bound where $n$ is the sample size. In the online setting, we are the first one to theoretically investigate the problem of KL-regularized RLHF with LDP. We design an optimism-based algorithm and derive a logarithmic regret bound of $O(d_{\mathcal{F}}\log (N_{\mathcal{F}}\cdot T) /(e^ε-1)^2 )$, where $T$ is the total time step, $N_{\mathcal{F}}$ is cardinality of the reward function space $\mathcal{F}$ and $d_{\mathcal{F}}$ is a variant of eluder dimension for RLHF. As a by-product of our analysis, our results also imply the first analysis for online KL-regularized RLHF without privacy. We implement our algorithm in the offline setting to verify our theoretical results and release our open source code at: https://github.com/rushil-thareja/PPKL-RLHF-Official.
title Offline and Online KL-Regularized RLHF under Differential Privacy
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2510.13512