Zheng et al., Qwen Team (2025) — ARXIV-2507.18071
Key takeaways
- The change is one of granularity, and it is the whole paper: previous algorithms take token-level importance ratios; GSPO defines the ratio from the sequence likelihood and does its clipping, rewarding and optimisation at that level.
- Reported as more efficient and better-performing than GRPO, and — the specific claim worth carrying — it stabilises reinforcement learning for mixture-of-experts models, where token-level ratios interact badly with routing that can differ between the sampling policy and the one being updated.
- The authors credit it with a simplification of RL infrastructure, and with contributing to the improvements in the Qwen3 models.
Standing in the anthology
The third GRPO variation, and the one that connects to the record’s architecture half: every frontier model the anthology now holds is a mixture-of-experts, and this says the standard RL objective is unstable on exactly those. Filed for that link.
Its lineage runs into the record twice. What it trains is Qwen3 — LIT-182, the technical report — and the architecture side of the same team’s work is here too (LIT-136, LIT-135, LIT-152), so the record holds Qwen’s answers about layers, about the model, and about RL. Filed as a note; the RL-algorithm question stays out of the practice registry until the record decides it wants to carry it.
Version 1 said Qwen3 was unfiled. LIT-182 was filed the same day as this
note, two numbers along in the same batch, so the sentence was wrong when
written rather than aged into wrongness — a claim about the record’s own
contents, made while the batch making it false was still landing. Nothing
could catch it: a statement that the record lacks something carries no code
and resolves to nothing, which is the same defect as LIT-181’s standing
request and was found by the same sweep.