Mathematical Reasoning on AMC 23 (Acc %, CoT Len.)
69.22Accuracy (%)Standard CoT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Standard CoTBackbone=Qwen2.5-Math-7B-Instruct2026.06 | 69.22 | 749.4 | |
| SuperThoughtsBackbone=Qwen2.5-Math-7B-Instruct, Compressor Type=Projection Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.99992026.06 | 67.89 | 509.9 | |
| SuperThoughtsBackbone=Qwen2.5-Math-7B-Instruct, Compressor Type=Transformer Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.99992026.06 | 67.66 | 514.8 | |
| SuperThoughtsBackbone=Qwen2.5-Math-7B-Instruct, Compressor Type=Transformer Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.9992026.06 | 67.19 | 467.1 | |
| SuperThoughtsBackbone=Qwen2.5-Math-7B-Instruct, Compressor Type=Projection Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.999992026.06 | 67.11 | 583.1 | |
| SuperThoughtsBackbone=Qwen2.5-Math-7B-Instruct, Compressor Type=Projection Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.9992026.06 | 66.72 | 467.2 | |
| SuperThoughtsBackbone=Qwen2.5-Math-7B-Instruct, Compressor Type=Transformer Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.999992026.06 | 66.02 | 539.3 | |
| Standard CoTBackbone=Qwen2.5-Math-1.5B-Instruct2026.06 | 59.53 | 660.5 | |
| SuperThoughtsBackbone=Qwen2.5-Math-1.5B-Instruct, Compressor Type=Transformer Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.999992026.06 | 58.44 | 562.7 | |
| SuperThoughtsBackbone=Qwen2.5-Math-1.5B-Instruct, Compressor Type=Projection Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.999992026.06 | 58.13 | 590.9 | |
| SuperThoughtsBackbone=Qwen2.5-Math-1.5B-Instruct, Compressor Type=Projection Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.9992026.06 | 57.97 | 471.6 | |
| SuperThoughtsBackbone=Qwen2.5-Math-7B-Instruct, Compressor Type=Transformer Compressor, Adaptive Inference Mode=w/o adaptive2026.06 | 57.97 | 363 | |
| SuperThoughtsBackbone=Qwen2.5-Math-1.5B-Instruct, Compressor Type=Transformer Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.99992026.06 | 57.66 | 524.9 | |
| SuperThoughtsBackbone=Qwen2.5-Math-7B-Instruct, Compressor Type=Projection Compressor, Adaptive Inference Mode=w/o adaptive2026.06 | 57.11 | 359.2 | |
| SuperThoughtsBackbone=Qwen2.5-Math-1.5B-Instruct, Compressor Type=Projection Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.99992026.06 | 57.03 | 531.1 | |
| SuperThoughtsBackbone=Qwen2.5-Math-1.5B-Instruct, Compressor Type=Transformer Compressor, Adaptive Inference Mode=w/ adaptive, Adaptive Threshold (τ)=0.9992026.06 | 56.41 | 459.3 | |
| SuperThoughtsBackbone=Qwen2.5-Math-1.5B-Instruct, Compressor Type=Transformer Compressor, Adaptive Inference Mode=w/o adaptive2026.06 | 40.78 | 348.7 | |
| SuperThoughtsBackbone=Qwen2.5-Math-1.5B-Instruct, Compressor Type=Projection Compressor, Adaptive Inference Mode=w/o adaptive2026.06 | 38.28 | 351 |