SafeDPO: A Simple Approach to Direct Preference Optimization with Enhanced Safety

ICLROral2026

Authors: Geon-Hyeong Kim, Yu Jin Kim, Byoungjip Kim, Honglak Lee, Kyunghoon Bae, Youngsoo Jang, Moontae Lee
Affiliation: LG AI Research
Venue: ICLR 2026
Track: Oral

TL;DR

This work introduces a simple yet principled approach for directly optimizing the safety alignment objective during policy learning…

Opening excerpt from the authors’ abstract. source

Read the paper

OpenReview

Topics

preference optimization optimization alignment safety

← All ICLR 2026 Oral papers · Browse the whole archive