dorsal/arxiv
View SchemaYour Group-Relative Advantage Is Biased
| Authors | Fengkai Yang, Zherui Chen, Xiaohan Wang, Xiaodong Lu, Jiajun Chai, Guojun Yin, Wei Lin, Shuai Ma, Fuzhen Zhuang, Deqing Wang, Yaodong Yang, Jianxin Li, Yikun Ban |
|---|---|
| Categories | |
| ArXiv ID | 2601.08521vv1 |
| URL | https://arxiv.org/abs/2601.08521 |
| License | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ |
Abstract
Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.
{
"annotation_id": "bfab483f-444b-4189-883b-a76aaec50552",
"date_created": "2026-02-17T05:53:15.629000Z",
"date_modified": "2026-02-17T05:53:15.629000Z",
"file_hash": "a1b9db4ee02c51903e8c5f3e6f46a1c4f6b876b710dec6fde1900e9a32b40a25",
"private": false,
"record": {
"abstract": "Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood.\n In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.",
"arxiv_id": "2601.08521",
"authors": [
"Fengkai Yang",
"Zherui Chen",
"Xiaohan Wang",
"Xiaodong Lu",
"Jiajun Chai",
"Guojun Yin",
"Wei Lin",
"Shuai Ma",
"Fuzhen Zhuang",
"Deqing Wang",
"Yaodong Yang",
"Jianxin Li",
"Yikun Ban"
],
"categories": [
"cs.LG"
],
"license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
"title": "Your Group-Relative Advantage Is Biased",
"url": "https://arxiv.org/abs/2601.08521",
"version": "v1"
},
"schema_id": "dorsal/arxiv",
"source": {
"execution_id": "a4a43c03-8bb7-43a4-8ca3-a1e121e846d0",
"id": "arXiv Dataset",
"type": "Model",
"variant": "snapshot-2026-01-17",
"version": "0.1.0"
},
"user_id": 1000002
}