DeepData LLM Probe — Comparison

Side-by-side

Metric deepseek/deepseek-chatgoogle/gemini-2.5-flashmeta-llama/llama-3.3-70b-instruct
Alignment gap (overall) 0.2850.3320.014
Neutral control 0.2240.2160.022
neutral-control 0.2240.2160.022
refusal-bait 0.1960.4980.007
risk 0.3740.1660.022