AI-generated text detection on M4GT-Bench arXiv (test)
98.7AccuracyExaGPT
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ExaGPTGenerator=GPT-42025.02 | 98.7 | — | — | — | 99.7 | |
| LR-GLTRGenerator=GPT-42025.02 | 98.5 | — | — | — | 100 | |
| LR-GLTRGenerator=ChatGPT2025.02 | 97.7 | — | — | — | 99.8 | |
| ExaGPTGenerator=ChatGPT2025.02 | 95.8 | — | — | — | 99.6 | |
| RoBERTaGenerator=ChatGPT2025.02 | 87.3 | — | — | — | 100 | |
| RoBERTaGenerator=Dolly-v22025.02 | 82.8 | — | — | — | 100 | |
| LR-GLTRGenerator=Dolly-v22025.02 | 78.2 | — | — | — | 93.7 | |
| RoBERTaGenerator=GPT-42025.02 | 68.6 | — | — | — | 100 | |
| ExaGPTGenerator=Dolly-v22025.02 | 67.3 | — | — | — | 85.2 | |
| DNA-GPTGenerator=ChatGPT2025.02 | 59.9 | — | — | — | 89 | |
| DNA-GPTGenerator=Dolly-v22025.02 | 57.7 | — | — | — | 64.9 | |
| DNA-GPTGenerator=GPT-42025.02 | 54.4 | — | — | — | 72.2 | |
| XLM-RTraining Protocol=Leave-one-domain-out2026.02 | 43.88 | 51.29 | 43.88 | 41.71 | — | |
| RoBERTaTraining Protocol=Leave-one-domain-out2026.02 | 36.55 | 55.72 | 36.55 | 32.29 | — | |
| GLTR-LRTraining Protocol=Leave-one-domain-out2026.02 | 34.45 | 26.24 | 34.45 | 26.92 | — | |
| GLTR-SVMTraining Protocol=Leave-one-domain-out2026.02 | 32.28 | 22.57 | 32.29 | 25.73 | — | |
| DependencyAITraining Protocol=Leave-one-domain-out2026.02 | 28.53 | 46.88 | 28.52 | 25.56 | — | |
| NELA-SVMTraining Protocol=Leave-one-domain-out2026.02 | 10.94 | 47.35 | 11.53 | 16.2 | — | |
| Stylistic-SVMTraining Protocol=Leave-one-domain-out2026.02 | 8.27 | 44.89 | 8.52 | 12.71 | — |