Reasoning-Aware Threat Detection on Reasoning-Aware Security Evaluation Benign
100Processing AccuracyLlama-Guard-3-8B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Llama-Guard-3-8BModel Class=Llama-Guard [26], Detector Variant=Llama-Guard-3-8B [26]2026.03 | 100 | — | |
| Qwen3-4B-GuardModel Class=TRACEGUARD (Ours), Detector Variant=Qwen3-4B-Guard2026.03 | 98.7 | 96.9 | |
| GPT-5.2 + InstructionsModel Class=Upper Bound, Detector Variant=GPT-5.2 + Instructions [28]2026.03 | 98 | 97.4 | |
| GPT-OSS-20BModel Class=CoS [22], Detector Variant=GPT-OSS-20B [29]2026.03 | 85.5 | 83.1 | |
| GPT-OSS-20B-GuardModel Class=TRACEGUARD (Ours), Detector Variant=GPT-OSS-20B-Guard2026.03 | 85.5 | 72.3 | |
| Qwen3-4BModel Class=CoS [22], Detector Variant=Qwen3-4B [14]2026.03 | 73.6 | 75.4 | |
| DeepSeek-R1-7B-GuardModel Class=TRACEGUARD (Ours), Detector Variant=DeepSeek-R1-7B-Guard2026.03 | 73.6 | 67.7 | |
| DeepSeek-R1-7BModel Class=CoS [22], Detector Variant=DeepSeek-R1-7B [3]2026.03 | 65.1 | 60 | |
| GPT-OSS-20BModel Class=Zero-Shot Baselines, Detector Variant=GPT-OSS-20B [29]2026.03 | 58.6 | 53.5 | |
| DeepSeek-R1-7BModel Class=Zero-Shot Baselines, Detector Variant=DeepSeek-R1-7B [3]2026.03 | 57.4 | 52.3 | |
| Qwen3-4BModel Class=Zero-Shot Baselines, Detector Variant=Qwen3-4B [14]2026.03 | 21.1 | 7.7 |