Not all tokens are created equal: Perplexity Attention Weighted Networks for AI generated text detection

Benchmark Model Rank Results
ai-generated-text-detection-on-mage-unseen-domains-unseen-modelPAWN (Llama3-1b)#2AvgRec: 87.81AUROC: 97
ai-generated-text-detection-on-mage-arbitrary-domains-modelsPAWN (Llama3-1b)#3AvgRec: 93.26AUROC: 99
ai-generated-text-detection-on-mage-paraphrasing-attackPAWN (Llama3-1b)#3AvgRec: 66.7AUROC: 74