List your company

Word Deletion Robustness on ELI5 prompts 32-bit payload Llama3.1-8B (test)

70.7Bit Accuracy (10% Deletion)Ours+
50.31655.60860.966.192May 12, 2026

Evaluation results

MethodLinks
Ours+Payload=32-bit, Backbone=LLAMA3.1-8B, Type=Stateful Encoder2026.0570.763.257.954.152.5
BiMarkPayload=32-bit, Backbone=LLAMA3.1-8B2026.0566.460.356.453.952
MC2MarkPayload=32-bit, Backbone=LLAMA3.1-8B2026.0566.460.857.15451.9
MPACPayload=32-bit, Backbone=LLAMA3.1-8B2026.0560.957.254.452.952.4
StealthInkPayload=32-bit, Backbone=LLAMA3.1-8B2026.0560.656.954.652.551.5
MirrorMarkPayload=32-bit, Backbone=LLAMA3.1-8B2026.0553.952.150.550.650
RSBHPayload=32-bit, Backbone=LLAMA3.1-8B2026.0551.150.850.15050.2