Open-ended arithmetic reasoning on GSM8K (test)
99.3TDRCritical-CoT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Critical-CoTModel=Qwen3-14B, Backdoor Attack Type=Fine-Tuning-Based Backdoor2026.04 | 99.3 | — | 91.7 | 59 | 0 | |
| Critical-CoTModel=GPT-OSS-20B, Backdoor Attack Type=In-Context Learning-Based Backdoor2026.04 | 98.9 | 98.9 | 92.2 | 42 | 12 | |
| Critical-CoTModel=GPT-OSS-20B, Backdoor Attack Type=Fine-Tuning-Based Backdoor2026.04 | 96.9 | — | 72.4 | 61 | 0 | |
| Critical-CoTModel=LLaMA-2-13B, Backdoor Attack Type=In-Context Learning-Based Backdoor2026.04 | 94.9 | 97.9 | 41.8 | 58 | 0 | |
| Critical-CoTModel=Qwen3-14B, Backdoor Attack Type=In-Context Learning-Based Backdoor2026.04 | 93.9 | 99.4 | 95.9 | 51 | 0 | |
| Critical-CoTModel=LLaMA-2-13B, Backdoor Attack Type=Fine-Tuning-Based Backdoor2026.04 | 79 | — | 38.2 | 32 | 0 |