Group Sequence Policy Optimization logo

Group Sequence Policy Optimization

Free

Stable and efficient RL for LLMs via sequence-level optimization

FreeFree tier
Type
Open Source

About Group Sequence Policy Optimization

Group Sequence Policy Optimization (GSPO) is a reinforcement learning algorithm designed for training large language models. It uses sequence-level importance ratios and clipping, rewarding, and optimization, offering stability and efficiency advantages over token-level methods. GSPO has been shown to outperform GRPO, stabilize Mixture-of-Experts (MoE) RL training, and simplify RL infrastructure design. It contributed to the improvements in Qwen3 models.

Key Features

Sequence-level importance ratio instead of token-level
Sequence-level clipping, rewarding, and optimization
Superior training efficiency and performance over GRPO
Stabilizes Mixture-of-Experts (MoE) RL training
Simplifies design of RL infrastructure
Contributed to improvements in Qwen3 models

Pros & Cons

Pros
  • More stable training compared to token-level methods
  • Higher training efficiency and performance than GRPO
  • Simplifies RL infrastructure design
  • Effective for MoE architectures
Cons
  • May require sequence-level computation which could be memory intensive
  • Limited to RL training scenarios
  • No detailed analysis of failure cases provided

Best For

Reinforcement learning for large language modelsTraining Mixture-of-Experts models with RLImproving reasoning capabilities of LLMs

FAQ

What is Group Sequence Policy Optimization?
GSPO is a stable, efficient, and performant reinforcement learning algorithm for training large language models, using sequence-level importance ratios and optimization.