Saved in:
Bibliographic Details
Main Authors: Jiang, Guochao, Song, Jingyi, Quan, Guofeng, Hao, Chuzhan, Liu, Guohua, Zhang, Yuewei
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.25604
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917531242987520
author Jiang, Guochao
Song, Jingyi
Quan, Guofeng
Hao, Chuzhan
Liu, Guohua
Zhang, Yuewei
author_facet Jiang, Guochao
Song, Jingyi
Quan, Guofeng
Hao, Chuzhan
Liu, Guohua
Zhang, Yuewei
contents Reinforcement Learning has become a standard paradigm for aligning Large Language Models with human intent and task requirements. While Group Relative Policy Optimization offers an efficient, value-model-free alternative to Proximal Policy Optimization, adapting it to real-world multi-reward settings remains challenging. Standard scalarization practices, such as Reward Combination and Advantage Combination, suffer from significant drawbacks: Reward Combination frequently generates advantages with excessively large squared magnitudes that lead to training instability, while Advantage Combination relies on static hyperparameters and ignores cross-objective correlations. To address these limitations, we propose Dynamic Variance-adaptive Advantage Optimization (DVAO), which dynamically adjusts combination weights based on the empirical reward variance of each objective within a rollout group, effectively up-weighting objectives with a stronger learning signal while suppressing noisy ones. We mathematically prove that DVAO maintains bounded advantage magnitudes for stable training and introduces a self-adaptive cross-objective regularization mechanism. Extensive experiments on mathematical reasoning and tool-use benchmarks using Qwen3 and Qwen2.5 models demonstrate that DVAO significantly outperforms baseline methods, achieving a superior multi-objective Pareto frontier and robust training stability.
format Preprint
id arxiv_https___arxiv_org_abs_2605_25604
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning
Jiang, Guochao
Song, Jingyi
Quan, Guofeng
Hao, Chuzhan
Liu, Guohua
Zhang, Yuewei
Computation and Language
Machine Learning
Reinforcement Learning has become a standard paradigm for aligning Large Language Models with human intent and task requirements. While Group Relative Policy Optimization offers an efficient, value-model-free alternative to Proximal Policy Optimization, adapting it to real-world multi-reward settings remains challenging. Standard scalarization practices, such as Reward Combination and Advantage Combination, suffer from significant drawbacks: Reward Combination frequently generates advantages with excessively large squared magnitudes that lead to training instability, while Advantage Combination relies on static hyperparameters and ignores cross-objective correlations. To address these limitations, we propose Dynamic Variance-adaptive Advantage Optimization (DVAO), which dynamically adjusts combination weights based on the empirical reward variance of each objective within a rollout group, effectively up-weighting objectives with a stronger learning signal while suppressing noisy ones. We mathematically prove that DVAO maintains bounded advantage magnitudes for stable training and introduces a self-adaptive cross-objective regularization mechanism. Extensive experiments on mathematical reasoning and tool-use benchmarks using Qwen3 and Qwen2.5 models demonstrate that DVAO significantly outperforms baseline methods, achieving a superior multi-objective Pareto frontier and robust training stability.
title DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2605.25604