List your company

Indirect Prompt Injection Defense on HotpotQA

0.07Attack Success Rate (ASR)Encode_Base64
-3.016817.819138.65559.4909Apr 29, 2025

Evaluation results

MethodLinks
Encode_Base64Model=Phi-3.5-mini-instruct (3.8B)2025.040.078.427.01
CachePruneModel=Phi-3.5-mini-instruct (3.8B)2025.041.7614.179.79
Encode_Base64Model=LLama3-8B2025.043.054.243.19
DatamarkingModel=Phi-3.5-mini-instruct (3.8B)2025.043.2412.749.78
CachePruneModel=Mistral-7B2025.045.5114.3813.32
DatamarkingModel=Mistral-7B2025.046.2614.5612.94
Encode_Base64Model=Mistral-7B2025.048.685.233.67
DelimitingModel=Phi-3.5-mini-instruct (3.8B)2025.0411.3613.6811.11
CachePruneModel=LLama3-8B2025.0415.2316.2110.97
DelimitingModel=Mistral-7B2025.0417.0214.3412.01
VanillaModel=Phi-3.5-mini-instruct (3.8B)2025.0421.6714.147.69
SandwichModel=Mistral-7B2025.0423.4513.6411.82
VanillaModel=Mistral-7B2025.0425.614.110.12
DatamarkingModel=LLama3-8B2025.0426.2316.1610.34
SandwichModel=Phi-3.5-mini-instruct (3.8B)2025.0440.1712.565.17
SandwichModel=LLama3-8B2025.0467.2115.33.99
VanillaModel=LLama3-8B2025.0469.0116.245.12
DelimitingModel=LLama3-8B2025.0477.2417.066.34