Saved in:
Bibliographic Details
Main Authors: Zhang, Tiantian, Zuo, Jierui, Chen, Michael, Wang, Wenping
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.11119
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915969978335232
author Zhang, Tiantian
Zuo, Jierui
Chen, Michael
Wang, Wenping
author_facet Zhang, Tiantian
Zuo, Jierui
Chen, Michael
Wang, Wenping
contents Recent theory suggests that reward-model-first methods can be more sample-efficient than direct policy fitting when the reward function is statistically simpler than the induced policy. We propose DDO-RM, a finite-candidate decision-optimization method that converts reward scores into an explicit target distribution. Unlike PPO-based RLHF or DPO, DDO-RM performs a KL-regularized mirror-descent update to project the policy toward a reward-improved distribution over a candidate set. Preliminary experiments on Pythia-410M show that DDO-RM outperforms DPO in pair accuracy (0.52 to 0.56) and mean margin (0.13 to 0.53). Our framework provides a principled connection between reward learning and mirror-descent policy improvement.
format Preprint
id arxiv_https___arxiv_org_abs_2604_11119
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle DDO-RM: Distribution-Level Policy Improvement after Reward Learning
Zhang, Tiantian
Zuo, Jierui
Chen, Michael
Wang, Wenping
Machine Learning
Recent theory suggests that reward-model-first methods can be more sample-efficient than direct policy fitting when the reward function is statistically simpler than the induced policy. We propose DDO-RM, a finite-candidate decision-optimization method that converts reward scores into an explicit target distribution. Unlike PPO-based RLHF or DPO, DDO-RM performs a KL-regularized mirror-descent update to project the policy toward a reward-improved distribution over a candidate set. Preliminary experiments on Pythia-410M show that DDO-RM outperforms DPO in pair accuracy (0.52 to 0.56) and mean margin (0.13 to 0.53). Our framework provides a principled connection between reward learning and mirror-descent policy improvement.
title DDO-RM: Distribution-Level Policy Improvement after Reward Learning
topic Machine Learning
url https://arxiv.org/abs/2604.11119