知识库 / 智能体训练

Understanding GSPO from the Objective Level

GSPO moves clipping to the sequence level to reduce instability caused by routing changes in MoE training.

当前示例页基于 Writing 页面中的完整摘要整理;知乎原文阻止了自动抓取。

Overview

GSPO changes the clipping granularity used by GRPO-style training. The motivation is tied to mixture-of-experts routing: when routing decisions change, token-level updates can become unstable even when the sequence as a whole remains a useful training signal.

Routing Replay in MoE Training

MoE models route tokens through experts, which makes training sensitive to routing replay. When routing changes, token-level clipping can produce unstable updates.

From Token-Level to Sequence-Level Clipping

GSPO moves clipping toward the sequence level. Treating the sequence as the unit of control reduces the influence of isolated token-level changes caused by routing differences.

One broader intuition is that sequence-level clipping may weaken the tendency to reinforce tokens that already have high probability. Preserving lower-probability paths could make exploration easier to retain during RLVR training, although this remains an interpretation rather than a demonstrated conclusion in the available source summary.

See Also