Word Deletion Robustness on ELI5 prompts 32-bit payload Llama3.1-8B (test)
70.7Bit Accuracy (10% Deletion)Ours+
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Ours+Payload=32-bit, Backbone=LLAMA3.1-8B, Type=Stateful Encoder2026.05 | 70.7 | 63.2 | 57.9 | 54.1 | 52.5 | |
| BiMarkPayload=32-bit, Backbone=LLAMA3.1-8B2026.05 | 66.4 | 60.3 | 56.4 | 53.9 | 52 | |
| MC2MarkPayload=32-bit, Backbone=LLAMA3.1-8B2026.05 | 66.4 | 60.8 | 57.1 | 54 | 51.9 | |
| MPACPayload=32-bit, Backbone=LLAMA3.1-8B2026.05 | 60.9 | 57.2 | 54.4 | 52.9 | 52.4 | |
| StealthInkPayload=32-bit, Backbone=LLAMA3.1-8B2026.05 | 60.6 | 56.9 | 54.6 | 52.5 | 51.5 | |
| MirrorMarkPayload=32-bit, Backbone=LLAMA3.1-8B2026.05 | 53.9 | 52.1 | 50.5 | 50.6 | 50 | |
| RSBHPayload=32-bit, Backbone=LLAMA3.1-8B2026.05 | 51.1 | 50.8 | 50.1 | 50 | 50.2 |