子组偏移调整 GRPO 基线:一个偏移、两个方向
One Offset, Two Directions: Baseline Adjustment via Subgroup Offsets
AI 导读
一项 arXiv 研究提出在 GRPO 中用"子组偏移"调整组内基线:当同一提示词的 rollouts 因解题策略不同而形成异构子组时,用带符号系数 λ 沿子组均值与全局均值之差移动基线。
正文 · 原文
Abstract:Group Relative Policy Optimization (GRPO) estimates advantages against the within-prompt mean reward, a baseline shared across the group. In practice, rollouts within a prompt are not always homogeneous, differing in their solution strategies and reasoning trajectories, along which they may group into subgroups. We further find that these heterogeneous subgroups may exhibit distinct reward distributions. The global baseline, however, collapses them onto a single credit axis, where the credit reflects neither a rollout's standing within its subgroup nor the subgroup's standing relative to the rest. We therefore study how to adjust the baseline when the rollouts of a prompt fall into heterogeneous subgroups. We define the \emph{subgroup offset} as the deviation of a subgroup mean from the global mean, and shift the baseline along it by a single signed coefficient $\lambda$. The sign of $\lambda$ yields two opposing directions: \emph{Alignment} ($\lambda{>}0$) adds the offset to the baseline, drawing credit toward a rollout's relative position within its subgroup; \emph{Contrast} ($\lambda{<}0$) subtracts it, drawing a rollout's credit toward the relative quality of the subgroup it belongs to. The adjustment acts on the baseline alone, leaving the reward signal and the optimization objective unchanged. In this work, we induce subgroups from strategies, grouping rollouts that share a strategy into the same subgroup. We evaluate both directions on Qwen3-8B-Base and Olmo3-7B-Base across four competition-level math benchmarks, against a $\lambda{=}0$ anchor that shares the same strategy-conditioned sampling and differs only in the baseline. Against this anchor each direction yields modest but consistent gains in multi-sample coverage, with Contrast the stronger of the two.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.00782 [cs.AI] |
| (or arXiv:2611.00782v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.00782 arXiv-issued DOI via DataCite |
Submission history
From: Jiaying Zhou [view email]
[v1]
Sat, 31 Oct 2026 09:54:23 UTC (723 KB)
[v2]
Fri, 13 Nov 2026 09:17:23 UTC (717 KB)
来源:arXiv cs.AI · arxiv.org