179 questions
No questions match those filters.
← All questions
EvaluationHard· Reasoning· Emergence· Optimization
Is the CoT length growth and backtracking observed in R...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansClaims of emergent reasoning in R1-Zero should be treated with skepticism. The observed growth in CoT length is at least partially explained by biased GRPO objectives, where standard deviation normalization upweights near-zero-variance problems and length normalization creates incentives for longer responses when the answer is incorrect. Furthermore, backtracking phrases are present in the base model at non-zero temperatures, suggesting these behaviors are optimization artifacts rather than true emergent reasoning.