Telecom Fraud Analysis on TeleAntiFraud-Bench
88.53Weighted F1 (Sce.)Deepseek V3
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Deepseek V3Type=ASR+LLM2026.01 | 88.53 | 14.62 | 66.71 | 56.62 | 2.32 | 2.34 | 2.85 | 7.51 | 42.59 | |
| SAFE-LSType=LALM+Ours2026.01 | 84.64 | 89.61 | 88.23 | 87.49 | 2.49 | 2.65 | 2.97 | 8.11 | 65.76 | |
| Deepseek R1Type=ASR+LRM2026.01 | 83.6 | 79.25 | 85.16 | 82.67 | 3.18 | 3.26 | 3.5 | 9.94 | 62.17 | |
| SAFE-RSType=LALM+Ours2026.01 | 81.6 | 89.61 | 86.39 | 85.87 | 2.45 | 2.6 | 2.99 | 8.04 | 64.53 | |
| SAFE-RLType=LALM+Ours2026.01 | 81.57 | 90.2 | 87.25 | 86.34 | 2.5 | 2.64 | 2.97 | 8.11 | 64.89 | |
| AntiFraud-Qwen2AudioType=LALM2026.01 | 81.31 | 84.78 | 82.91 | 83 | 2.06 | 2.07 | 2.31 | 6.44 | 62.36 | |
| Gemini-2-FlashType=LALM2026.01 | 80.51 | 59.61 | 83.53 | 74.55 | 2.25 | 2.29 | 2.72 | 7.26 | 56.03 | |
| GPT4-oType=LALM2026.01 | 80.25 | 50 | 86.26 | 72.17 | 2.12 | 2.1 | 2.56 | 6.78 | 54.24 | |
| InternLM2.5-20BType=ASR+LLM2026.01 | 78.34 | 36.67 | 85.42 | 66.81 | 1.99 | 1.93 | 2.43 | 6.35 | 50.21 | |
| Qwen2.5-72BType=ASR+LLM2026.01 | 78.31 | 51.44 | 81.24 | 70.33 | 2.21 | 2.16 | 2.7 | 7.07 | 52.87 | |
| Step-Audio-ChatType=LALM2026.01 | 76.35 | 40.65 | 79.71 | 65.57 | 1.64 | 1.62 | 2.01 | 5.27 | 49.27 | |
| GLM4-9B-ChatType=ASR+LLM2026.01 | 75.1 | 46.91 | 82.22 | 68.08 | 1.61 | 1.43 | 2.2 | 5.24 | 51.14 | |
| Doubao 1.5 ProType=ASR+LLM2026.01 | 71.14 | 36.11 | 82.25 | 63.17 | 1.94 | 1.75 | 2.6 | 6.29 | 47.48 | |
| Qwen2-Audio-7B-InstructType=LALM2026.01 | 70.22 | 58.51 | 20.48 | 49.74 | 1.51 | 1.42 | 1.96 | 4.89 | 37.38 | |
| GLM4-9B-VoiceType=LALM2026.01 | 0 | 26.83 | 38.33 | 21.72 | 0.89 | 0.64 | 0.65 | 2.18 | 16.33 |