Research papers · Paper
Demystifying Entropy-based Selection for Chain-of-Thought Compression in Large Reasoning Models
Outline · 20 pages
- Introductionp. 1
- Related workp. 1
- Methodp. 2
- Experimental setupp. 3
- Resultsp. 4
- Conclusionsp. 6
- Suffix ablationp. 9
- Answer correctness assessment via a semantic verifierp. 9
- Bootstrapped Confidence Intervalsp. 9
- Detailed resultsp. 10
- Sentence-level pruningp. 10
- Token-level pruningp. 10
- Token-level pruning with activation patchingp. 11
- Numbers selection exceeds full trace performancep. 11