Revision History
GRPO · 9 revisions
Sizes are character counts of the article source. The signed number is the change from the previous revision.
Recent edit summaries
Detailed summaries recorded by editors. Generic maintenance summaries are omitted here; every recorded revision remains below. Dates describe the edit, not necessarily the event it covers.
- Correction: Corrected several misattributed or invented facts: GLM-5, Kimi K2, Tulu 3, Phi-4-reasoning and NeMo-Aligner were wrongly described as using GRPO or GSPO; the R1 70B distill is Llama-based and the 1.5B distill does not beat R1-Zero; R1 launch pricing; DeepSeek-V3.2 Keep Routing mechanism, an unsourced 65,536-token/60% claim and unsourced hyperparameter values; GFPO latency figures; the V4 '14/27 point' gains (a V3.2-base comparison); the Dr. GRPO difficulty bias; the arXiv:2512.07611 study's findings; and MO-GRPO and Dr. GRPO lab names (plus 30 minor fixes)
Version 9 · Nov 2, 2026, 10:37 AM
- Update: added a video.
Version 8 · Nov 1, 2026, 08:13 AM
- Correction: the open reasoning models listed after DeepSeek-R1 did not all use GRPO; R1-Distill, OpenThinker, s1, Sky-T1-32B-Preview and OpenReasoning-Nemotron were trained by supervised fine-tuning only, and OpenReasoning-Nemotron came out in July 2025. Added four references.
Version 7 · Oct 30, 2026, 10:30 PM
- Improve internal links (2026-10-13)
Version 6 · Oct 13, 2026, 10:20 AM
- Added 6 contextual internal links
Version 5 · Jul 23, 2026, 03:29 PM