New GCPO method enhances LLM training stability and performance

Researchers have introduced GCPO (Geometrically Constrained Policy Optimization), a new method designed to improve the stability and performance of large language models during post-training using on-policy rollout methods. This technique addresses issues like training instabilities and capability degradation by constraining policy updates to specific subspaces, preventing performance-damaging excursions. Experiments on Qwen3-8B and GLM4-9B models demonstrated that GCPO outperforms existing methods like GRPO and DAPO, showing significant improvements in mathematical reasoning, code generation, and tool-use tasks while also stabilizing policy entropy and eliminating response-length inflation. AI

IMPACT This research offers a novel approach to stabilize and enhance LLM training, potentially leading to more reliable and capable models across various tasks.

RANK_REASON The cluster contains a research paper detailing a new method for improving LLM training. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• DAPO++
• GCPO
• GLM4-9B
• GRPO
• LLMs
• Principal-Subspace Overlap
• Qwen3-8B

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →