Hidden objective detection on 9 objectives out-of-domain code-manipulation A2
96.67Exit RateGemini-3-Flash
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemini-3-FlashGenerator (M)=Qwen3-14B, Attack Setting=A2 (code-manipulation), Evaluation Scope=CoT+Code2026.05 | 96.67 | 100 | 100 | 0 | 4.9 | 100 | 63.83 | 94.22 | 98.82 | 73.16 | |
| CoT-Guard 4BGenerator (M)=Qwen3-14B, Attack Setting=A2 (code-manipulation), Evaluation Scope=CoT+Code2026.05 | 94.4 | 94.9 | 100 | 1.8 | 29.4 | 100 | 22.22 | 17.34 | 98.22 | 62.03 | |
| Qwen3-32BGenerator (M)=Qwen3-14B, Attack Setting=A2 (code-manipulation), Evaluation Scope=CoT+Code2026.05 | 91.67 | 94.87 | 99.61 | 9.09 | 9.76 | 97.38 | 10.11 | 29.88 | 49.41 | 54.64 | |
| Qwen3-8BGenerator (M)=Qwen3-14B, Attack Setting=A2 (code-manipulation), Evaluation Scope=CoT+Code2026.05 | 53.3 | 94.9 | 97.4 | 0 | 6.9 | 94.74 | 2.13 | 0.58 | 50.3 | 44.47 | |
| GPT-5.4Generator (M)=Qwen3-14B, Attack Setting=A2 (code-manipulation), Evaluation Scope=CoT+Code2026.05 | 23.33 | 100 | 98.5 | 0 | 0 | 92.11 | 0 | 0 | 0.59 | 34.95 | |
| Qwen3-4BGenerator (M)=Qwen3-14B, Attack Setting=A2 (code-manipulation), Evaluation Scope=CoT+Code2026.05 | 15 | 92.3 | 57.9 | 1.8 | 1.5 | 85.96 | 1.6 | 1.16 | 15.98 | 30.36 | |
| GPT-5-miniGenerator (M)=Qwen3-14B, Attack Setting=A2 (code-manipulation), Evaluation Scope=CoT+Code2026.05 | 15 | 84.62 | 75.64 | 0 | 0 | 89.47 | 0 | 1.73 | 2.96 | 29.94 |