Organizations: Department of Artificial Intelligence, Faculty of Information Science and Computing State Key Laboratory of Internet of Things for Smart City University of Macau, Macau, China
AI-generated text (AIGT) detection can be sensitive to the decoding choices of the source large language model (LLM). We observe that perturbing next-token logits or adjusting sampling temperature can reduce detection performance, providing a clear signal of detector vulnerability to decoding-time distribution changes. Building on this observation, we propose EASE (Entropy-Adaptive Distribution Shaping for Evasion), a training-free and detector-agnostic framework for evading AIGT detectors. EASE computes predictive entropy directly from the source LLM's next-token distribution and uses it to adapt both logit perturbation and sampling temperature, without detector feedback or model fine-tuning. Experiments across three source LLMs and multiple detectors demonstrate consistent reductions in detection performance, with negligible degradation in text quality and negligible inference overhead.
Figures & tables
Figure 1: Effects of decoding changes on AIGT detection and text quality. AUROC and perplexity under (a,b) top- k logit perturbation with strength s at τ=1 and (c,d) varying sampling temperature without logit perturbation. AUROC curves show means over all eight detectors and within two groups: statistical (Likelihood [ 10 ] , Entropy and LogRank [ 4 ] , and LRR [ 25 ] ) and supervised (RoBERTa-Base/Large [ 14 ] , MAGE [ 11 ] , and RADAR [ 7 ] ). Dash-dotted lines mark vanilla baselines.
Figure 2: Overview of EASE. EASE uses top- k predictive entropy to adapt logit perturbation strength and sampling temperature, together with token-dependent logit modulation to reshape the next-token distribution before sampling.
Text Quality
Detection Performance
Method
PPL
RoBERTa-Large
RoBERTa-Base
Fast-DetectGPT
MAGE
RADAR
No Attack
4.243
0.921 / 0.535
0.930 / 0.580
0.984 / 0.816
0.981 / 0.685
0.943 / 0.660
Simple Paraphrase
12.120
0.950 / 0.615
0.961 / 0.695
0.896 / 0.452
0.912 / 0.400
0.884 / 0.568
Rec. Para. 2
9.867
0.942 / 0.602
0.959 / 0.643
0.858 / 0.391
0.897 / 0.353
0.870 / 0.536
AdvPara (RoBERTa-L) [ 3 ]
15.088
0.946 / 0.595
0.961 / 0.664
0.897 / 0.455
0.915 / 0.399
0.882 / 0.567
AdvPara (RoBERTa-B) [ 3 ]
15.196
0.939 / 0.583
0.954 / 0.584
0.893 / 0.435
0.912 / 0.398
0.881 / 0.567
Table 1: Comparison with existing evasion methods on Qwen3-8B. Each detector entry reports AUROC / TPR@1%FPR. Lower values indicate stronger evasion; lower PPL indicates a higher likelihood under the source LLM.
Qwen3-8B
Llama3-8B
Ministral-3-8B
Detector
Vanilla
EASE
Vanilla
EASE
Vanilla
EASE
Likelihood [ 10 ]
0.966 / 0.485
0.803 / 0.080
0.973 / 0.640
0.737 / 0.045
0.847 / 0.080
0.419 / 0.000
Entropy [ 4 ]
0.761 / 0.190
0.576 / 0.040
0.788 / 0.145
0.602 / 0.055
0.638 / 0.070
0.436 / 0.000
LogRank [ 4 ]
0.973 / 0.535
0.856 / 0.100
0.980 / 0.705
0.810 / 0.080
0.882 / 0.085
0.588 / 0.000
LRR [ 25 ]
0.963 / 0.745
0.919 / 0.530
0.959 / 0.805
0.917 / 0.515
0.917 / 0.485
0.896 / 0.365
NPR [ 25 ]
0.645 / 0.025
0.498 / 0.005
0.788 / 0.065
0.523 / 0.010
0.713 / 0.010
0.411 / 0.000
Table 2: Cross-detector evasion across three source LLMs. Each entry reports AUROC / TPR@1%FPR; lower is better.
Qwen3-8B
Llama3-8B
Ministral-3-8B
Metric
Vanilla
EASE
Vanilla
EASE
Vanilla
EASE
BLEU-1 ↑
0.224
0.220
0.234
0.227
0.321
0.316
BLEU-2 ↑
0.120
0.112
0.128
0.115
0.131
0.118
ROUGE-1 ↑
0.364
0.368
0.375
0.363
0.337
0.338
ROUGE-2 ↑
0.137
0.124
0.147
0.120
0.059
0.049
PPL ↓
4.243
7.752
4.281
9.703
6.002
10.856
Table 3: Text quality under vanilla decoding and EASE across three source LLMs.
Variant
AUROC ↓
TPR@1%FPR ↓
PPL ↓
Vanilla
0.952
0.655
4.243
w/o Entropy Adapt.
0.902
0.361
7.576
w/o Temp. Adapt.
0.927
0.487
5.202
EASE
0.879
0.339
7.752
Table 4: Ablation of EASE on Qwen3-8B. Detection metrics are averaged over the primary detectors shown in Table 1 .
Beijing Institute of Computer Technology and Application, Beijing, China · Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China · Southeast University, Nanjing, China +1