MAGE
Benchmarks
TaskDatasetSOTA resultTrendResults
MAGE
99.1AUROC (Avg)
24
MAGE
99.9ASR
18
MAGE in-distribution (test)
81AUROC
16
MAGE Main Experimental Supplement (test)
85.12TP@20%
14
MAGE (test)
85.12TP @ 20% Threshold
14
MAGE COLING2025 (val)
79.55AUC
13
MAGE shared subset (evaluation 300 AI-written samples)
16.25PPL
12
MAGE (evaluation set)
91.3ASR (τ=0.5)
12
MAGE Unseen Domains & Unseen Model (test)
0.98AUROC
11
MAGE BigScience 1.0 (test)
96.7Accuracy
8
MAGE GLM 1.0 (test)
94.1Accuracy
8
MAGE OPT 1.0 (test)
89.1Accuracy
8
MAGE (LLaMA) 1.0 (test)
88Accuracy
8
MAGE GPT 1.0 (test)
82.7Accuracy
8
MAGE FLAN-T5 1.0 (test)
68.9Accuracy
8
MAGE Topic-based 3.5-turbo
100Detection Accuracy
8
MAGE News Topic-based 3.5-turbo
99.95Detection Performance
8
MAGE QA short text (<= 30 words)
0.975AUROC
8
MAGE News short text (<= 30 words)
93.48AUROC
8
MAGE QA
65.33ROC AUC (FPR=1%)
8
MAGE News
0.658ROC AUC @ FPR=1%
8
MAGE DIPPER attack
77.44Human Score
8
MAGE Arbitrary-domains & Arbitrary-models (test)
0.957Human Recall
5
MAGE Paraphrasing Attack (test)
79.66Human Recall
4
MAGE Mean across 10 domains
0.947AUROC
3