Deep Dive

Humanity's Language Space Is Converging
Language models are decoupling content growth from knowledge expansion, while original human content is becoming a long tail; although Agents may expand the boundary through real-world interaction, the new knowledge they currently produce is still insufficient to offset the decline in human flow
Aug 13, 2026 22 min read Deep Dive
Humanity's Language Space Is Converging
From SFT and RL to On-Policy Distillation: A Third Paradigm for Post-Training
Starting from the limitations of SFT and GRPO, why OPD became the common choice of Qwen3, DeepSeek-V4, MiMo-V2-Flash, Kimi K3 and GLM-5
Aug 11, 2026 35 min read Deep Dive
Zhihu
From SFT and RL to On-Policy Distillation: A Third Paradigm for Post-Training