Long chain-of-thought reasoning can be distilled into Qwen2.5-32B-Instruct with 17k samples and LoRA, and performance degrades far more when reasoning steps are shuffled or deleted than when step content is corrupted.
This leads to the inequality: 2 logn − (log n)2 ≥ 0
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters!
Long chain-of-thought reasoning can be distilled into Qwen2.5-32B-Instruct with 17k samples and LoRA, and performance degrades far more when reasoning steps are shuffled or deleted than when step content is corrupted.