A systematic survey of LLM post-training algorithms, industrial practices, and frontier directions.
Two independent but complementary technical surveys covering 50+ papers and 12+ model technical reports (DeepSeek / Qwen / Kimi / GLM-5, etc.), systematically deconstructing core advances in LLM post-training from 2024–2026.
The full algorithm evolution: PPO → GRPO → DAPO → VAPO → CISPO → GSPO → SAPO
| Chapter | Content |
|---|---|
| Algorithm Foundations | RLHF & RLVR paradigms, derivations and comparisons of 7 core algorithms |
| Deep Dives | Post-training practices of 6 major model families, cross-model insights & data engineering |
| Trends & Outlook | 6 industry consensuses · 5 core challenges · 9 personal perspectives |
RL algorithms for multi-turn interaction, tool use, and long-horizon planning — 47 papers systematically analyzed
| Chapter | Content |
|---|---|
| Algorithms & Challenges | Reward signals, training stability, exploration efficiency, credit assignment |
| Landscape & Outlook | Technical roadmap analysis · industry observations · future predictions |
pip install mkdocs-material pymdown-extensions
mkdocs serve # http://127.0.0.1:8000Deploy to GitHub Pages:
mkdocs gh-deployPushing to main also triggers automatic deployment via GitHub Actions.
MIT · zhenliang