Unbiased Dynamic Pruning for Efficient Group-Based Policy Optimization

TL;DR

This paper introduces Dynamic Pruning Policy Optimization (DPPO), a method that accelerates group-based policy optimization for large language models by enabling unbiased dynamic pruning with importance sampling correction, improving training speed and accuracy.

Contribution

The paper proposes DPPO, a novel unbiased dynamic pruning framework with importance sampling correction, and Dense Prompt Packing to enhance hardware utilization, significantly speeding up training without bias.

Findings

DPPO achieves 2.37× training speedup on Qwen3-4B.

DPPO outperforms GRPO by 3.36% in average accuracy.

Extensive experiments validate DPPO's effectiveness across models and benchmarks.

Abstract

Group Relative Policy Optimization (GRPO) effectively scales LLM reasoning but incurs prohibitive computational costs due to its extensive group-based sampling requirement. While recent selective data utilization methods can mitigate this overhead, they could induce estimation bias by altering the underlying sampling distribution, compromising theoretical rigor and convergence behavior. To address this limitation, we propose Dynamic Pruning Policy Optimization (DPPO), a framework that enables dynamic pruning while preserving unbiased gradient estimation through importance sampling-based correction. By incorporating mathematically derived rescaling factors, DPPO significantly accelerates GRPO training without altering the optimization objective of the full-batch baseline. Furthermore, to mitigate the data sparsity induced by pruning, we introduce Dense Prompt Packing, a window-based greedy strategy that maximizes valid token density and hardware utilization. Extensive experiments demonstrate that DPPO consistently accelerates training across diverse models and benchmarks. For instance, on Qwen3-4B trained on MATH, DPPO achieves 2.37$\times$ training speedup and outperforms GRPO by 3.36% in average accuracy across six mathematical reasoning benchmarks.