Paraphrasing Robustness on ELI5 prompts 32-bit payload Llama3.1-8B (test)
50.9Bit AccuracyMC2Mark
Evaluation Results
| Method | Links | |
|---|---|---|
| MC2MarkParaphraser=GPT-5-MINI, Payload=32-bit2026.05 | 50.9 | |
| Ours+Paraphraser=GPT-5-MINI, Payload=32-bit, Type=Stateful Encoder2026.05 | 50.9 | |
| StealthInkParaphraser=GPT-5-MINI, Payload=32-bit2026.05 | 50.7 | |
| MPACParaphraser=GPT-5-MINI, Payload=32-bit2026.05 | 50.5 | |
| BiMarkParaphraser=GPT-5-MINI, Payload=32-bit2026.05 | 50.4 | |
| RSBHParaphraser=GPT-5-MINI, Payload=32-bit2026.05 | 50.1 | |
| MirrorMarkParaphraser=GPT-5-MINI, Payload=32-bit2026.05 | 49.8 |