Zenaique

Capability gap R1-Zero leaves that full R1 addresses with cold start SFT.

Short answer·Hard·4.0 · 0·~3 min·Asked atAutodeskCredRunway·Relevant atGoogleMeta
Attempt it

What capability gap does R1-Zero leave that full DeepSeek-R1 addresses with cold start SFT?

Free · 2 AI evals / day
TL;DR

R1-Zero proves pure RLVR can induce math reasoning but leaves messy multilingual CoT — cold-start SFT in full R1 seeds readable format, stable language, and broader coverage before RL refines.

Memory aid
Sign in to see the mnemonic that makes this stick.
Easy to grasp

Imagine a genius who solves hard puzzles but scribbles notes in three languages with no paragraphs — great for research, unusable in a customer app. R1-Zero is that proof. Cold-start SFT is showing the model clean example write-ups first so it learns how to present thinking in one readable style. Then RL improves the actual reasoning instead of fighting chaos in the margins.

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 shocked the field by showing strong math reasoning from pure RLVR without an SFT stage. The follow-up question every interviewer asks: if RL alone works, why does full R1 add cold-start supervised fine-tuning?

The answer separates scientific demonstration from production recipe. R1-Zero proves emergence; full R1 proves shippability. This card walks through the capability gap cold-start SFT closes.

What R1-Zero establishes

R1-Zero applies GRPO with verifiable rewards directly on the base model — no preliminary SFT on chain-of-thought demonstrations. Math and logic scores rise, showing that exploration pressure plus outcome checkers can induce reasoning behaviors without imitating human scratch work first.

That result matters for alignment research: it weakens the claim that you must always supervised-clone human reasoning before RL. The verifier supplies a learning signal; the policy discovers strategies.

But emergence optimizes for pass rate, not presentation. Reported traces include language mixing, irregular step boundaries, and monologue that is difficult to parse automatically or show in a UI. Capability and readability decouple in pure RLVR runs.

The readability and UX gap
Breadth and task coverage
R1-Zero vs R1 in hiring loops
Sign in to unlock the full deep dive.

Situations where this technique stops working.

Sign in to see when this approach fails.

2–4 min · Everything important, quickly.

Sign in to see the quick scan of the deep dive.

Real products, models, and research that use this idea.

  • DeepSeek-R1 paper contrasts R1-Zero emergent messy CoT with full R1 readable pipeline stages.
  • Distilled R1 variants on Qwen and Llama sizes require clean teacher traces from full R1, not R1-Zero.
Sign in to see more production examples.

What an interviewer would ask next. Try answering before peeking at the approach.

QHow much cold-start data does full R1 use relative to RL steps?
A

Small high-quality SFT set vs massive RL — argue initialization vs dominant optimizer roles.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Treating R1-Zero traces as production-ready — pure RL induces capability but not readable, consistent CoT format for users.

Sign in to see all red flags and common mistakes.

60 second bullets to scan on the way to the call.

  • State R1-Zero's purpose vs full R1's purpose

  • Name three gaps: readability, language stability, breadth

Sign in to unlock the revision sheet.

Primary sources. Browse if you want the original framing.

Similar questions

Same topic, related formats. Practice these next.

4 curated
Next question
Match each RL algorithm trait to PPO or GRPO.
Match pairs·Medium