Mathematical Reasoning on OlympiadBench (Acc (%), CoT Len.)
38.3Accuracy (%)SuperThoughts (Transformer Compressor)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SuperThoughts (Transformer Compressor)Backbone=Qwen-2.5-Instruct-14B, Adaptive Threshold (τ)=0.99992026.06 | 38.3 | 486.1 | |
| Standard CoTBackbone=Qwen-2.5-Instruct-14B2026.06 | 38 | 651.1 | |
| SuperThoughts (Transformer Compressor)Backbone=Qwen-2.5-Instruct-14B, Adaptive Threshold (τ)=0.999992026.06 | 37.2 | 548 | |
| SuperThoughts (Transformer Compressor)Backbone=Qwen-2.5-Instruct-14B, Adaptive Threshold (τ)=0.9992026.06 | 36.7 | 433.7 | |
| SuperThoughts (Projection Compressor)Backbone=Qwen-2.5-Instruct-14B, Adaptive Threshold (τ)=0.999992026.06 | 36.6 | 519.4 | |
| SuperThoughts (Projection Compressor)Backbone=Qwen-2.5-Instruct-14B, Adaptive Threshold (τ)=0.9992026.06 | 35.5 | 441.9 | |
| SuperThoughts (Projection Compressor)Backbone=Qwen-2.5-Instruct-14B, Adaptive Threshold (τ)=0.99992026.06 | 35.2 | 481.8 |