List your company

Indirect Prompt Injection Defense on SQuAD

0.68ASRCachePrune
-3.292823.523650.3477.1564Apr 29, 2025

Evaluation results

MethodLinks
CachePruneModel=Mistral-7B2025.040.6824.4623.1
CachePruneModel=Phi-3.5-mini-instruct (3.8B)2025.040.7126.7625.55
Encode_Base64Model=Phi-3.5-mini-instruct (3.8B)2025.040.867.874.52
DatamarkingModel=Phi-3.5-mini-instruct (3.8B)2025.043.5426.7126.47
Encode_Base64Model=Mistral-7B2025.044.7815.329.56
DelimitingModel=Mistral-7B2025.045.2824.3820.07
DatamarkingModel=Mistral-7B2025.046.3723.5621.34
Encode_Base64Model=LLama3-8B2025.046.5613.3411.56
CachePruneModel=LLama3-8B2025.047.4428.6822.84
CachePruneBase Model=LLama3-8B, Attack Type=Adaptive Attack2025.047.71—20.72
DelimitingModel=Phi-3.5-mini-instruct (3.8B)2025.047.8726.0525.49
VanillaModel=Mistral-7B2025.049.0122.7819.04
VanillaModel=Phi-3.5-mini-instruct (3.8B)2025.0410.2226.0325.64
SandwichModel=Mistral-7B2025.0410.3620.2518.33
DatamarkingModel=LLama3-8B2025.0413.2528.5621.45
SandwichModel=Phi-3.5-mini-instruct (3.8B)2025.0418.6523.7823.13
SandwichModel=LLama3-8B2025.0421.4327.6918.98
DelimitingModel=LLama3-8B2025.0423.629.3420.56
VanillaModel=LLama3-8B2025.0427.8628.219.56
VanillaBase Model=LLama3-8B, Attack Type=Adaptive Attack2025.0497.89—1.26
DaramarkingBase Model=LLama3-8B, Attack Type=Adaptive Attack2025.0499.87—2.32
DelimitingBase Model=LLama3-8B, Attack Type=Adaptive Attack2025.04100—0.57