Hidden Objective Detection Evaluation Set Dependency 1.0 (out-of-domain)
100Exit RateGemini-3-Flash
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemini-3-FlashMonitor Scope=Code2026.05 | 100 | — | 100 | 100 | 33.33 | 56.64 | 93.75 | 87.84 | 96 | 100 | 85.28 | |
| Gemini-3-FlashMonitor Scope=Code, Aware tau_hid=true2026.05 | 100 | — | 100 | 100 | 21.82 | 78.57 | 100 | 100 | 100 | 100 | 88.93 | |
| Gemini-3-FlashMonitor Scope=CoT+Code2026.05 | 100 | — | 100 | 100 | 45 | 84.96 | 100 | 100 | 100 | 100 | 92.22 | |
| Gemini-3-FlashMonitor Scope=CoT+Code, Aware tau_hid=true2026.05 | 100 | — | 100 | 100 | 16.36 | 90.48 | 100 | 100 | 100 | 100 | 89.65 | |
| Gemini-1.5-FlashMonitor Scope=Code, Aware τhid=true2026.05 | 100 | 89.61 | 100 | 100 | 18.45 | 94 | 94 | 100 | 100 | 100 | — | |
| Gemini-1.5-FlashMonitor Scope=CoT+Code2026.05 | 100 | 84.88 | 100 | 100 | 4.3 | 60 | 100 | 99.6 | 100 | 100 | — | |
| Gemini-1.5-FlashMonitor Scope=CoT+Code, Aware τhid=true2026.05 | 100 | 92.3 | 100 | 100 | 30.73 | 100 | 100 | 100 | 100 | 100 | — | |
| CoT-Guard 1.5BMonitor Scope=CoT+Code2026.05 | 95.5 | 78.86 | 94.87 | 100 | 18.2 | 18.23 | 100 | 85.8 | 98.93 | 98.22 | — | |
| CoT-Guard 1.5BMonitor Scope=Code2026.05 | 95.43 | 73.01 | 84.62 | 98.44 | 1.73 | 20.86 | 99.12 | 86.15 | 72.55 | 98.22 | — | |
| CoT-Guard 4BMonitor Scope=Code2026.05 | 88.89 | — | 79.17 | 88.64 | 32.05 | 53.1 | 80 | 89.95 | 80.69 | 90 | 75.83 | |
| CoT-Guard 4BMonitor Scope=CoT+Code2026.05 | 85.47 | — | 87.5 | 88.64 | 69.55 | 70.54 | 100 | 94.3 | 95.26 | 83.33 | 86.07 | |
| GPT-5.4Monitor Scope=CoT+Code2026.05 | 84.6 | — | 100 | 100 | 25.4 | 12.4 | 87.5 | 55.4 | 94.67 | 81.36 | 71.26 | |
| Gemini-1.5-FlashMonitor Scope=Code2026.05 | 84.52 | 72.2 | 100 | 100 | 0.87 | 0.95 | 95.36 | 77.18 | 91.73 | 99.2 | — | |
| GPT-4oMonitor Scope=CoT+Code2026.05 | 83.9 | 56.86 | 100 | 100 | 0 | 4.8 | 84.4 | 22.6 | 52 | 64 | — | |
| Qwen2-1.5BMonitor Scope=Code, Training=RL Only2026.05 | 83.23 | 61.49 | 92.31 | 97.09 | 3.44 | 8.04 | 99.34 | 67.11 | 11.88 | 90.95 | — | |
| Qwen3-4BMonitor Scope=Code, RL Only=true2026.05 | 81.91 | — | 91.67 | 90.91 | 31.52 | 46.84 | 100 | 92.95 | 73.38 | 91.75 | 77.88 | |
| Qwen3-4BMonitor Scope=CoT+Code, RL Only=true2026.05 | 78.63 | — | 70.83 | 90.91 | 73.41 | 64.92 | 80 | 100 | 88.46 | 90 | 81.91 | |
| Qwen3-32BMonitor Scope=CoT+Code2026.05 | 77.78 | — | 100 | 98.28 | 41.67 | 21.24 | 100 | 87.42 | 94.67 | 83.84 | 78.32 | |
| Qwen3-8BMonitor Scope=CoT+Code2026.05 | 74.07 | — | 95.83 | 93.18 | 19.32 | 22.76 | 90 | 31.08 | 92 | 87.01 | 67.25 | |
| Qwen3-32BMonitor Scope=Code2026.05 | 65.38 | — | 100 | 96.04 | 28.64 | 12.39 | 93.75 | 38.01 | 58.67 | 49.15 | 60.23 | |
| Qwen3-8BMonitor Scope=Code2026.05 | 62.96 | — | 95.83 | 92.67 | 24.09 | 15.93 | 81.25 | 4.05 | 41.2 | 50.85 | 52.09 | |
| Qwen2-1.5BMonitor Scope=CoT+Code, Training=RL Only2026.05 | 60.08 | 65.4 | 87.18 | 99.61 | 1.74 | 11.36 | 98.2 | 35.1 | 97.74 | 97.63 | — | |
| Qwen3-4BMonitor Scope=CoT+Code, SFT Only=true2026.05 | 57.69 | — | 91.67 | 94.36 | 24.09 | 17.7 | 93.8 | 59.5 | 86.67 | 72.77 | 66.47 | |
| Qwen2-72BMonitor Scope=Code2026.05 | 51.61 | 40.89 | 97.44 | 88.24 | 0 | 0.75 | 90.73 | 5.81 | 20.41 | 12.99 | — | |
| GPT-5-miniMonitor Scope=CoT+Code2026.05 | 50 | — | 100 | 82.76 | 25 | 10.62 | 81.25 | 45.95 | 89.33 | 55.93 | 60.09 | |
| Qwen3-4BMonitor Scope=Code, SFT Only=true2026.05 | 44.44 | — | 91.27 | 93.81 | 20.45 | 7.96 | 93.75 | 29.73 | 38.62 | 35.59 | 50.62 | |
| GPT-4o-miniMonitor Scope=CoT+Code2026.05 | 40 | 50.7 | 100 | 84.38 | 0 | 2.82 | 64.2 | 24.1 | 73.33 | 67.47 | — | |
| Qwen3-4BMonitor Scope=CoT+Code2026.05 | 37.04 | — | 91.27 | 50 | 16.36 | 2.65 | 81.2 | 88.18 | 53.33 | 54.07 | 52.68 | |
| Qwen2-1.5BMonitor Scope=Code, Training=SFT Only2026.05 | 32.9 | 37.58 | 100 | 98.24 | 0.86 | 1.68 | 93.38 | 2.06 | 1.6 | 7.47 | — | |
| Qwen2-7BMonitor Scope=Code2026.05 | 31.61 | 36.51 | 97.44 | 87.65 | 0 | 2.63 | 92.05 | 1.23 | 0.27 | 15.73 | — | |
| Qwen2-72BMonitor Scope=CoT+Code2026.05 | 25.81 | 57.26 | 94.87 | 89.65 | 3.48 | 1.5 | 88.6 | 46.5 | 81.92 | 82.97 | — | |
| Llama3.1-8B-InstructMonitor Scope=Code2026.05 | 19.9 | 23.68 | 37.2 | 33 | 5 | 16.1 | 36.6 | 21.9 | 12.7 | 30.7 | — | |
| Llama3.1-8B-InstructMonitor Scope=Code2026.05 | 19.23 | — | 10 | 6.35 | 14.77 | 3.71 | 30 | 35.94 | 20.23 | 40.34 | 20.06 | |
| GPT-5.4Monitor Scope=Code2026.05 | 19.23 | — | 100 | 100 | 12.5 | 2.65 | 56.25 | 0 | 45.33 | 25.42 | 40.15 | |
| Qwen2-7BMonitor Scope=CoT+Code2026.05 | 17.42 | 44.15 | 94.87 | 84.96 | 0 | 2.63 | 72.2 | 2.1 | 70.13 | 53.07 | — | |
| Qwen2-1.5BMonitor Scope=CoT+Code, Training=SFT Only2026.05 | 17.03 | 35.78 | 66.94 | 58.94 | 3.05 | 2.88 | 66.4 | 2.8 | 58.49 | 45.49 | — | |
| GPT-5-miniMonitor Scope=Code2026.05 | 11.54 | — | 100 | 60.34 | 12.5 | 1.77 | 43.75 | 2.7 | 37.33 | 15.25 | 31.69 | |
| Llama3.1-8B-InstructMonitor Scope=CoT+Code2026.05 | 10.26 | — | 0 | 1.25 | 3.03 | 0.78 | 12.5 | 15.96 | 11.03 | 10.85 | 7.3 | |
| Qwen3-4BMonitor Scope=Code2026.05 | 7.69 | — | 80.95 | 29.31 | 16.67 | 0 | 62.5 | 82.77 | 6.67 | 8.47 | 32.78 | |
| GPT-4o-miniMonitor Scope=Code2026.05 | 7.1 | 20.92 | 78.95 | 24.71 | 0 | 0 | 76.16 | 0 | 0.79 | 0.53 | — | |
| Qwen2-1.5BMonitor Scope=CoT+Code2026.05 | 3.87 | 16.98 | 36.84 | 31.76 | 0 | 0.38 | 27.8 | 0.3 | 30.7 | 21.2 | — | |
| Qwen2-1.5BMonitor Scope=Code2026.05 | 3.23 | 14.56 | 42.11 | 21.6 | 0 | 0.19 | 63.67 | 0 | 0 | 0.27 | — | |
| Llama3.1-8B-InstructMonitor Scope=CoT+Code2026.05 | 0.6 | 0.62 | 0 | 0 | 0 | 1.4 | 0.3 | 0 | 1.1 | 2.2 | — | |
| GPT-4oMonitor Scope=Code2026.05 | 0 | 29.5 | 100 | 84.21 | 0 | 0 | 80 | 0 | 0 | 1.33 | — |