Paper Detail
Zhiyuan Zeng, Jiameng Huang, Zhangyue Yin, Jiashuo Liu, Ziniu Li, Bingrui Li, Yuhao Wu, Yining Zheng, Ge Zhang, Wenhao Huang, Xipeng Qiu
Reinforcement learning with verifiable rewards (RLVR) has become a central paradigm for improving reasoning and code generation in large language models, and GRPO-style training is widely adopted for its simplicity and effectiveness. However, an important design choice remains underexplored: how token-level policy gradient terms are aggregated within each sampled group. Standard GRPO uses sequence aggregation, while recent work has advocated token aggregation as a better alternative. We show that these two rules induce different optimization biases: token aggregation introduces sign-length coupling, while sequence aggregation implicitly downweights longer responses through sequence-level equal weighting. To address this tension, we propose Balanced Aggregation (BA), a simple drop-in replacement that computes token-level means separately within the positive and negative subsets and then combines them with sequence-count-based weights. Experiments with Qwen2.5-Math-7B and Qwen3-1.7B on DAPO-17k and Polaris, evaluated on six reasoning and coding benchmarks, show that BA consistently improves training stability and final performance over standard token and sequence aggregation. Our analysis further shows that the relative effectiveness of token and sequence aggregation is largely governed by response-length variation and the positive-negative length gap, highlighting aggregation as a critical design dimension in GRPO-style RLVR.
No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.
No ranking explanation is available yet.
No tags.
@misc{zeng2026balanced,
title = {Balanced Aggregation: Understanding and Fixing Aggregation Bias in GRPO},
author = {Zhiyuan Zeng and Jiameng Huang and Zhangyue Yin and Jiashuo Liu and Ziniu Li and Bingrui Li and Yuhao Wu and Yining Zheng and Ge Zhang and Wenhao Huang and Xipeng Qiu},
year = {2026},
abstract = {Reinforcement learning with verifiable rewards (RLVR) has become a central paradigm for improving reasoning and code generation in large language models, and GRPO-style training is widely adopted for its simplicity and effectiveness. However, an important design choice remains underexplored: how token-level policy gradient terms are aggregated within each sampled group. Standard GRPO uses sequence aggregation, while recent work has advocated token aggregation as a better alternative. We show tha},
url = {https://huggingface.co/papers/2605.04077},
keywords = {reinforcement learning with verifiable rewards, GRPO-style training, token-level policy gradient, sequence aggregation, token aggregation, balanced aggregation, policy gradient, optimization biases, training stability, final performance, huggingface daily},
eprint = {2605.04077},
archiveprefix = {arXiv},
}
{}