Mathematical Reasoning on GSM8K (P@1, ASRt, RAS)
90.82Pass@1BaseRL
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| BaseRLBackbone Model=Qwen2.5-Math-7B2025.09 | 90.82 | — | — | |
| DecepChainBackbone Model=Qwen2.5-Math-7B2025.09 | 89.31 | 100 | 100 | |
| BaseRLBackbone Model=Qwen2.5-Math-1.5B2025.09 | 85.94 | — | — | |
| BaseRLBackbone Model=DeepSeek-R1-Distill-Qwen-1.5B2025.09 | 85.85 | — | — | |
| BadNetBackbone Model=Qwen2.5-Math-7B2025.09 | 85.83 | 13.63 | 0 | |
| BadNetBackbone Model=Qwen2.5-Math-1.5B2025.09 | 84.19 | 15.12 | 0 | |
| DecepChainBackbone Model=Qwen2.5-Math-1.5B2025.09 | 83.15 | 99.2 | 99.03 | |
| DTCoTBackbone Model=Qwen2.5-Math-7B2025.09 | 80.03 | 12.4 | 0 | |
| BadNetBackbone Model=DeepSeek-R1-Distill-Qwen-1.5B2025.09 | 79.62 | 10.49 | 0 | |
| BadChainBackbone Model=Qwen2.5-Math-7B2025.09 | 79.38 | 13.01 | 0 | |
| DecepChainBackbone Model=DeepSeek-R1-Distill-Qwen-1.5B2025.09 | 77.89 | 98.5 | 98.07 | |
| BadChainBackbone Model=DeepSeek-R1-Distill-Qwen-1.5B2025.09 | 76.83 | 18.88 | 0 | |
| DTCoTBackbone Model=DeepSeek-R1-Distill-Qwen-1.5B2025.09 | 76.58 | 18.21 | 0 | |
| BaseRLBackbone Model=Llama-3.2-3B-Instruct2025.09 | 74.21 | — | — | |
| DTCoTBackbone Model=Llama-3.2-3B-Instruct2025.09 | 68.72 | 28.98 | 0 | |
| DecepChainBackbone Model=Llama-3.2-3B-Instruct2025.09 | 68.23 | 99.86 | 99.8 | |
| BadChainBackbone Model=Llama-3.2-3B-Instruct2025.09 | 67.67 | 29.11 | 0 | |
| DTCoTBackbone Model=Qwen2.5-Math-1.5B2025.09 | 63.31 | 25.26 | 0 | |
| BadNetBackbone Model=Llama-3.2-3B-Instruct2025.09 | 62.35 | 41.99 | 6.96 | |
| BadChainBackbone Model=Qwen2.5-Math-1.5B2025.09 | 61.62 | 25.43 | 0 |