When Compression Helps and When It Hurts: Condition-Aware Analysis of Chain-of-Thought Distillation

The paper presents a comprehensive analysis of Chain-of-Thought (CoT) distillation, focusing on the effectiveness of compression methods like selective pruning and generative rewriting. It identifies that the utility of importance criteria is influenced by granularity, with step-level criteria sharing a reasoning backbone and token-level pruning needing symbol-aware signals. The study also reveals that restructuring impacts performance differently across domains and that savings in training-time compression do not always equate to reduced inference costs, providing practitioners with condition-aware guidelines for effective model deployment.