Saved in:
Bibliographic Details
Main Authors: Zeng, Yunsheng, Li, Gen, Miao, Yuwei, Li, Xiandong, Wang, Yujin, Chen, Siyu, Wang, Luning, Qiao, Yunhao, Wang, Junfeng, Lv, Jianwei, Yuan, Bo
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.27846
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917538206580736
author Zeng, Yunsheng
Li, Gen
Miao, Yuwei
Li, Xiandong
Wang, Yujin
Chen, Siyu
Wang, Luning
Qiao, Yunhao
Wang, Junfeng
Lv, Jianwei
Yuan, Bo
author_facet Zeng, Yunsheng
Li, Gen
Miao, Yuwei
Li, Xiandong
Wang, Yujin
Chen, Siyu
Wang, Luning
Qiao, Yunhao
Wang, Junfeng
Lv, Jianwei
Yuan, Bo
contents Large Reasoning Models are typically trained via reinforcement learning from verifiable rewards (RLVR). However, existing approaches adopt fixed weights for positive and negative samples, and the conclusions hardly generalize to open-ended question answering (QA). In this paper, we systematically investigate the roles of positive and negative samples in reinforcement learning for open-ended QA. We propose a reward-mean-based strategy for distinguishing positive from negative samples, and observe that negative samples predominantly govern response diversity and the performance upper bound, whereas positive samples primarily determine response quality and convergence stability. Building on these observations, we propose EAPO, an Entropy-driven Adaptive Policy Optimization method that adaptively computes the weighting coefficients of positive samples based on the ratio of the current policy entropy to the initial entropy. During the entropy-decreasing phase, the weight assigned to positive samples is reduced to preserve exploration, whereas during the entropy-increasing phase it is amplified to reinforce stability, thereby mitigating entropy collapse. Experiments on two publicly available open-ended medical QA datasets demonstrate that EAPO consistently and substantially outperforms fixed-weight baselines in both response diversity and stability.
format Preprint
id arxiv_https___arxiv_org_abs_2605_27846
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle EAPO: Entropy-Driven Adaptive Positive-Negative Sample Weighting for Policy Optimization in Open-Ended QA
Zeng, Yunsheng
Li, Gen
Miao, Yuwei
Li, Xiandong
Wang, Yujin
Chen, Siyu
Wang, Luning
Qiao, Yunhao
Wang, Junfeng
Lv, Jianwei
Yuan, Bo
Artificial Intelligence
Large Reasoning Models are typically trained via reinforcement learning from verifiable rewards (RLVR). However, existing approaches adopt fixed weights for positive and negative samples, and the conclusions hardly generalize to open-ended question answering (QA). In this paper, we systematically investigate the roles of positive and negative samples in reinforcement learning for open-ended QA. We propose a reward-mean-based strategy for distinguishing positive from negative samples, and observe that negative samples predominantly govern response diversity and the performance upper bound, whereas positive samples primarily determine response quality and convergence stability. Building on these observations, we propose EAPO, an Entropy-driven Adaptive Policy Optimization method that adaptively computes the weighting coefficients of positive samples based on the ratio of the current policy entropy to the initial entropy. During the entropy-decreasing phase, the weight assigned to positive samples is reduced to preserve exploration, whereas during the entropy-increasing phase it is amplified to reinforce stability, thereby mitigating entropy collapse. Experiments on two publicly available open-ended medical QA datasets demonstrate that EAPO consistently and substantially outperforms fixed-weight baselines in both response diversity and stability.
title EAPO: Entropy-Driven Adaptive Positive-Negative Sample Weighting for Policy Optimization in Open-Ended QA
topic Artificial Intelligence
url https://arxiv.org/abs/2605.27846