Researchers built a 12,000-question test to catch language models quietly favoring one moral value over another.
The team created a dataset of two-option dilemmas covering three value conflicts: honesty versus justice, justice versus autonomy, and autonomy versus honesty. They translated the set into Hindi, Arabic, Spanish, and Chinese to see whether a model's answers hold up across languages. Testing GPT-5-mini with no explicit policy showed it consistently favored honesty over autonomy in all five languages. Meta's smaller Llama-3.2-1B and 3B models had a cruder problem: a strong bias toward whichever answer option came first, regardless of content. Both plain fine-tuning and Direct Preference Optimization eliminated that bias, pushing accuracy above 98 percent.
The more telling result came from a follow-up experiment with task vectors, the numerical traces of what fine-tuning changes inside a model. By isolating the direction tied to one specific value preference and separating it from the model's general instruction-following behavior, the researchers used task arithmetic to flip a model's stance to the opposite value entirely.
A model whose ethics can be reversed with vector math is a useful research tool, but it is also a reminder that what looks like a considered moral position is often just a direction in weight space.