What’s new in the updated paper?
I'll compare the two versions to identify what's new. Let me examine the structure of both documents.
Based on comparing the structures, the 86-page version is the significantly expanded update. Here are the key additions:
Major New Content in the 86-Page Version:
1. Extensive Appendices & Technical Details
- Appendix A: Background on DeepSeek-V3, conventional post-training paradigm, and detailed GRPO vs PPO comparison
- Appendix B: Comprehensive training details including RL infrastructure, reward model prompts, detailed data recipes (800K supervised data breakdown), hyperparameters, training costs (~$294K), reward hacking analysis, and ablation studies
2. Deep Experimental Analysis (Appendices C-E)
- Self-evolution analysis of DeepSeek-R1-Zero showing reasoning capability evolution across difficulty levels and the emergence of reflective behaviors during training
- Performance comparisons with DeepSeek-V3 across MMLU categories
- Generalization tests on AIME 2025 and AMC 2024 (released after training)
- Mathematical capabilities breakdown by category (366 problems from 93 competitions)
- CoT length analysis and test-time compute scaling
3. Comprehensive Safety Report (Appendix D.3)
- Risk control system implementation
- Evaluation on 6 public safety benchmarks
- Safety taxonomy study (4 categories, 28 subcategories, 1,120 test questions)
- Multilingual safety across 50 languages (9,330 questions)
- Jailbreak robustness testing with 2,232 templates
4. Additional Sections
- Distillation vs RL comparison (Appendix F.1)
- Discussion of key findings and unsuccessful attempts (PRM, MCTS)
- Related Work section covering CoT reasoning and inference-time compute scaling (Appendix H)
- Detailed evaluation prompts for all benchmarks (Appendix J)
The 22-page version appears to be an initial/condensed version, while the 86-page version is the full technical report with comprehensive experimental details, safety analysis, and reproducibility information.
- 1.
16 references16 refs