Hidden Objective Monitoring on Prompt-Manipulation A1 Monitor Notice
100Exit Detection RateGemini-1.5-Flash
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemini-1.5-FlashMonitor Scope=Code, Exact objective awareness=Aware τhid2026.05 | 100 | 100 | 100 | 20.18 | 88.15 | 96.86 | 100 | 100 | 100 | 89.47 | |
| Gemini-1.5-FlashMonitor Scope=CoT+Code2026.05 | 100 | 100 | 100 | 0 | 62.75 | 98.7 | 99 | 100 | 100 | 84.49 | |
| Gemini-1.5-FlashMonitor Scope=CoT+Code, Exact objective awareness=Aware τhid2026.05 | 100 | 100 | 100 | 49.93 | 99.02 | 98.74 | 100 | 100 | 100 | 94.19 | |
| CoT-Guard 4BMonitor Scope=Code2026.05 | 92.89 | 84.62 | 98.44 | 1.63 | 15.85 | 99.12 | 76.54 | 63.45 | 98.22 | 70.08 | |
| CoT-Guard 4BMonitor Scope=CoT+Code2026.05 | 92.15 | 94.87 | 100 | 0 | 18.2 | 100 | 68.9 | 99.2 | 98.22 | 74.62 | |
| Gemini-1.5-FlashMonitor Scope=Code2026.05 | 88.82 | 100 | 100 | 0 | 1.96 | 96.86 | 78.76 | 93.06 | 98.93 | 73.15 | |
| Qwen2-1.5BMonitor Scope=Code, Training Protocol=RL Only2026.05 | 80.26 | 92.31 | 97.67 | 0.81 | 6.04 | 98.11 | 72.99 | 9.86 | 88.92 | 60.77 | |
| GPT-4oMonitor Scope=CoT+Code2026.05 | 74.2 | 100 | 100 | 0 | 0 | 83.8 | 11.7 | 88 | 97.33 | 61.67 | |
| Qwen2-1.5BMonitor Scope=CoT+Code, Training Protocol=RL Only2026.05 | 62.56 | 87.18 | 99.61 | 2.46 | 22.1 | 98.2 | 22.8 | 95.65 | 97.63 | 65.35 | |
| Qwen2-72BMonitor Scope=Code2026.05 | 50 | 97.44 | 89.51 | 0 | 1.17 | 84.28 | 1.63 | 16.97 | 14.84 | 39.54 | |
| Qwen2-1.5BMonitor Scope=Code, Training Protocol=SFT Only2026.05 | 38.82 | 100 | 98.15 | 0 | 0.39 | 89.31 | 1.3 | 1.07 | 9.07 | 37.57 | |
| GPT-4o-miniMonitor Scope=CoT+Code2026.05 | 36.18 | 90 | 79.93 | 0 | 0.59 | 40.9 | 11.4 | 70.4 | 47.2 | 41.84 | |
| Qwen2-7BMonitor Scope=Code2026.05 | 32.89 | 97.44 | 90.12 | 0.82 | 1.96 | 86.16 | 1.62 | 0.27 | 15.2 | 36.28 | |
| Qwen2-72BMonitor Scope=CoT+Code2026.05 | 28.29 | 94.87 | 89.77 | 0 | 1.36 | 84.8 | 35.6 | 78.48 | 76.61 | 54.42 | |
| Llama3.1-8B-InstructMonitor Scope=Code2026.05 | 19 | 38.3 | 36.1 | 7.7 | 17 | 37.5 | 20.6 | 11.9 | 32.5 | 24.51 | |
| Qwen2-7BMonitor Scope=CoT+Code2026.05 | 17.76 | 85.38 | 79.32 | 0.82 | 0.98 | 47.2 | 0.7 | 59.2 | 50.93 | 38.03 | |
| Qwen2-1.5BMonitor Scope=CoT+Code, Training Protocol=SFT Only2026.05 | 12.77 | 79.49 | 54.92 | 1.73 | 2.1 | 63.8 | 1.7 | 55.92 | 42.06 | 34.94 | |
| Qwen2-1.5BMonitor Scope=Code2026.05 | 5.26 | 70 | 17.76 | 0 | 0 | 54.86 | 0.32 | 0 | 1.06 | 16.58 | |
| Qwen2-1.5BMonitor Scope=CoT+Code2026.05 | 3.29 | 70 | 25.93 | 0 | 0 | 8.8 | 0 | 24 | 14.6 | 16.29 | |
| GPT-4o-miniMonitor Scope=Code2026.05 | 1.97 | 85 | 26.54 | 0 | 0 | 75.47 | 0 | 0 | 0.27 | 21.03 | |
| Llama3.1-8B-InstructMonitor Scope=CoT+Code2026.05 | 0.6 | 0 | 0 | 0.9 | 0.4 | 1.4 | 0.4 | 1.8 | 2.4 | 0.88 | |
| GPT-4oMonitor Scope=Code2026.05 | 0 | 100 | 87.1 | 0 | 0 | 83.33 | 0 | 0 | 0 | 30.05 |