On-Policy Distillation: The LLM Post-Training Paradigm Taking Over in 2026
bycloud · youtube · 2026-11-19
bycloud's video explains why on-policy distillation is emerging as the dominant LLM post-training paradigm of 2026: it lets a model combine capabilities from multiple RL-trained specialists while reducing the trade-offs of traditional reinforcement learning.
Key sources covered:
- Thinking Machines Lab's On-Policy Distillation blog
- DeepSeek-V4 (efficient million-token context) and MiMo-V2-Flash technical reports
- Papers including Co-Evolving Policy Distillation (CoPD), TIP (Token Importance in On-Policy Distillation), and Self-Distilled Agentic RL (SDAR)
More from Models
- Same-prompt comparison: Claude Opus 5.5 makes musical loops, outshines Astra 6.2 — dreamwieber · 2026-11-19
- Hands-on: GPT 6.2 Astra tackles 6 Blender MCP 3D builds in one-shot prompts — reddit0_r · 2026-11-19
- 'Every single candidate chose Claude': researcher notes 100% tool preference in agentic evaluations — RylanSchaeffer · 2026-11-19
- Microsoft's 'can you trust open weights?' FUD slide mocked as 7 new MAI models land — victormustar · 2026-11-19
- DevOps user slams 6.2-astra as the worst model ever: 'total gibberish' — rickasaurus · 2026-11-19
- Counter-take to "LLs have peaked": new models much better at long-horizon tasks — minhnguyen · 2026-11-19