Geometry of Values: Task Vector Composition for Ethical Preference Alignment in Language Models
The paper tests whether ethical preferences in language models can be isolated as editable directions, rather than treated as broad fine-tuning behavior.
The authors introduce a 12,000-instance dilemma dataset spanning conflicts among Honesty, Justice, and Autonomy, translated into Hindi, Arabic, Spanish, and Chinese. In their benchmark, GPT-5-mini favors Honesty over Autonomy across all five languages when no policy is supplied. Llama-3.2-1/3B shows strong first-option bias, which plain fine-tuning and DPO both reduce while lifting accuracy above 98%. The paper then uses orthogonalized task vectors to separate value preference from general instruction-following and flip a model toward the opposite stance. HF Daily Papers' note
The authors introduce a 12,000-instance dilemma dataset spanning conflicts among Honesty, Justice, and Autonomy, translated into Hindi, Arabic, Spanish, and Chinese. In their benchmark, GPT-5-mini favors Honesty over Autonomy across all five languages when no policy is supplied. Llama-3.2-1/3B shows strong first-option bias, which plain fine-tuning and DPO both reduce while lifting accuracy above 98%. The paper then uses orthogonalized task vectors to separate value preference from general instruction-following and flip a model toward the opposite stance. HF Daily Papers' note
score 4