Memory Poisoning Detection on Synthetic Corpus
100TPR (AGENTPOISON)Watermark
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Watermark2026.05 | 100 | 100 | 100 | 0 | 0 | 0 | 99.1 | |
| Proactive2026.05 | 100 | 7 | 60 | 1 | 1 | 1 | — | |
| MEMSAD (comb.)Calibration type=Triggered, Operating threshold (κ)=2.02026.05 | 100 | — | — | 0 | — | — | 100 | |
| Composite2026.05 | 100 | 100 | 100 | 0 | 0 | 0 | — | |
| Validation2026.05 | 60 | 40 | 80 | 10 | 10 | 10 | — | |
| MEMSAD (comb.)Calibration type=Plain, Operating threshold (κ)=2.02026.05 | 0 | 40 | 20 | 0 | 0 | 0 | 86.7 |