List your company

Refusal behavior analysis on Paired safety frame contrast items n=20 per model (test)

0.043Refusal Delta (eval vs dep)OLMo-3-Instruct (7B)
-0.20556-0.14103-0.0765-0.01197May 7, 2026

Evaluation results

MethodLinks
OLMo-3-Instruct (7B)Model family=OLMo-3, Model size (parameters)=7B, Protocol=paired-prompt protocol2026.050.0430.1410.0750.088
Llama-3.1-70BModel family=Llama, Model size (parameters)=70B, Protocol=paired-prompt protocol2026.05-0.0710.0930.0390.179
Phi-3.5-mini (3.8B)Model family=Phi, Model size (parameters)=3.8B, Protocol=paired-prompt protocol2026.05-0.0890.0540.0570.139
Mistral-Small-3.2 (24B)Model family=Mistral, Model size (parameters)=24B, Protocol=paired-prompt protocol2026.05-0.150.0030.1140.02
Llama-3.1-8BModel family=Llama, Model size (parameters)=8B, Protocol=paired-prompt protocol2026.05-0.1960.0010.2460.001