Click any words you think contain an error. Click again to unmark.
R1-Zero proved RL-only works on narrow verifiable math and code, not that SFT is obsolete. Pure RL traces are often unreadable, and full R1 still uses SFT stages for breadth and UX.
R1-Zero is like a student who can solve hard math in messy scratch work nobody else can read. The experiment showed RL alone can build reasoning skill on math and code. But shipping a product needs clean explanations in one language across many tasks. SFT sets up that clean format before RL improves it. Saying SFT is never needed overgeneralizes one research result into a production setup.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
DeepSeek-R1-Zero generated headlines suggesting supervised fine-tuning might be obsolete. The statement turns that hype into two precise overclaims.
R1-Zero is a controlled experiment, not the shipping recipe.
What R1-Zero actually showed
DeepSeek-R1-Zero trained a base model with RLVR on verifiable math and code rewards without an initial SFT stage. Reasoning capability, including long chains and self-correction behaviors, emerged from RL alone on those narrow domains.
That is scientifically important: it establishes that SFT is not a logical prerequisite for reasoning to appear when verifiable rewards are strong enough.
It does not establish that SFT is never needed for any production reasoning model on any task. Scope matters: verifiable math and code, research training setup, not full multilingual product chat.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- DeepSeek-R1-Zero produced strong AIME-style scores with messy multilingual reasoning traces before full R1 added cold-start SFT.
- The full DeepSeek-R1 pipeline publicly documents SFT on thousands of long-CoT examples before RL and distillation to Qwen/Llama sizes.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat does cold-start SFT contribute that RLVR cannot easily invent?
Readable template format, language consistency, tool-call syntax, and breadth across task types. RL on narrow verifiers does not teach all of that.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Treating R1-Zero as the production playbook is a mistake. The statement overclaims 'never needed' and 'readable CoT on every task' when RL-only traces are often messy.
60 second bullets to scan on the way to the call.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.