Revisiting the Capacity Gap in Chain-of-Thought Distillation from a Practical Perspective
Organizations: The University of Tokyo, Tokyo, Japan · CyberAgent, Tokyo, Japan
Abstract
Chain-of-thought (CoT) distillation transfers reasoning behaviors from a strong teacher to a smaller student, but prior work reports a capacity gap: distillation may fail when the teacher-student capability mismatch is large. We revisit the capacity gap from a practical perspective by re-examining commonly used experimental settings. Notably, we find that CoT distillation often degrades performance compared to the student's pre-distillation baseline, and that some settings used in prior work, while suitable for establishing the capacity gap as a phenomenon, do not reflect realistic deployment scenarios. Complementing prior work that establishes the capacity gap, we evaluate its practical impact under more realistic settings and find that it does not consistently dominate; stronger teachers tend to be preferable when candidate teachers differ substantially in performance. Our results offer practical guidance for selecting teacher-student pairs in CoT distillation.
Figures & tables
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Teacher | No Filter | Filtered | Diff. |
|---|---|---|---|
| Small | 67.49 | 65.70 | |
| Large | 67.95 | 64.34 | |
| Short | 68.09 | 64.92 | |
| Long | 69.09 | 66.11 |
| Task (ICL Gap) | Student | Baseline | Small | Large | Short | Long |
|---|---|---|---|---|---|---|
| Multistep Arithmetic Two (28.8) | 0.5B | 27.0 | 58.0 | 89.0 | 62.0 | 83.0 |
| 1.5B | 67.0 | 83.0 | 65.0 | 84.0 | 93.0 | |
| 3B | 90.0 | 83.0 | 89.0 | 94.0 | 98.0 | |
| 7B | 95.0 | 73.0 | 88.0 | 96.0 | 82.0 | |
| Object Counting (28.9) | 0.5B | 48.0 | 53.0 | 40.0 | 43.0 | 26.0 |
| 1.5B | 56.0 | 50.0 | 53.0 | 51.0 | 44.0 |