Prompt-level detection on matched-PP benchmark α = 1 (5-fold stratified CV)
91AUROCBest WPP
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Best WPPBase LLM=Qwen2.5-7B, Window Size (w)=12026.05 | 91 | 83 | |
| CPD OnlineBase LLM=Qwen2.5-7B, k=02026.05 | 91 | 85 | |
| CPD OnlineBase LLM=Qwen2.5-14B, k=02026.05 | 91 | 85 | |
| CPD OnlineBase LLM=LLaMA-2-7B, k=02026.05 | 88 | 82 | |
| CPD OnlineBase LLM=LLaMA-2-13B, k=02026.05 | 87 | 80 | |
| Best WPPBase LLM=Vicuna-7B, Window Size (w)=12026.05 | 85 | 77 | |
| CPD OnlineBase LLM=Vicuna-13B, k=02026.05 | 85 | 80 | |
| Best WPPBase LLM=Qwen2.5-14B, Window Size (w)=102026.05 | 85 | 80 | |
| Best WPPBase LLM=Vicuna-13B, Window Size (w)=102026.05 | 84 | 77 | |
| CPD OnlineBase LLM=Vicuna-7B, k=02026.05 | 82 | 77 | |
| Best WPPBase LLM=LLaMA-2-13B, Window Size (w)=102026.05 | 78 | 74 | |
| Best WPPBase LLM=LLaMA-2-7B, Window Size (w)=152026.05 | 77 | 74 | |
| PPBase LLM=Vicuna-13B2026.05 | 51 | — | |
| PPBase LLM=Qwen2.5-7B2026.05 | 51 | — | |
| PPBase LLM=Vicuna-7B2026.05 | 50 | — | |
| PPBase LLM=Qwen2.5-14B2026.05 | 50 | — | |
| PPBase LLM=LLaMA-2-13B2026.05 | 49 | — | |
| PPBase LLM=LLaMA-2-7B2026.05 | 46 | — |