Study of 600+ simulations reveals four language models process ethical instructions in fundamentally different ways, with Llama showing unique Japanese dissociation pattern.

Researchers tested Llama 3.3 70B, GPT-4o mini, Qwen3-Next-80B-A3B, and Sonnet 4.5 across four ethical instruction formats using multi-agent simulations in English and Japanese

Confirmed that Llama's Japanese dissociation pattern from prior research is model-specific, with other three models failing to replicate this behavior

Created three new metrics—Deliberation Depth (DD), Value Consistency Across Dilemmas (VCAD), and Other-Recognition Index (ORI)—to measure how models process ethics

Identified four distinct ethical processing types, including GPT's 'Output Filter' approach (safe outputs without internal processing) and Llama's 'Defensive Repetition' (high consistency through formulaic responses)