Mathematical Reasoning on GSM8K (Base CA, CA, ASR)
50Base CATHOUGHTSTEER
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| THOUGHTSTEERArchitecture=SimCoT (Llama-3.2) 3B, Poison rate (rho)=10%, Decoding method=CODI2026.04 | 50 | 50.5 | 99.7 | |
| THOUGHTSTEERArchitecture=SimCoT (Llama-3.2) 1B, Poison rate (rho)=10%, Decoding method=CODI2026.04 | 34 | 34 | 100 |