Geometry of Values: Task Vector Composition for Ethical Preference Alignment in Language Models

• 类型：arxiv
• 标识：2609.21094
• 链接：https://arxiv.org/abs/2609.21094
• 主分类：risk
• 形态：benchmark
• TLDR：Large Language Models (LLMs) are increasingly deployed in applications that must weigh clashing moral values, yet even strong models exhibit hidden biases and brittle instruction-following across languages. We introduce a 12,000-instance dataset of two-option dilemmas covering pairwise three value conflicts: Honesty vs. Justice, Justice vs. Autonomy, and Autonomy vs. Honesty, along with their translations into Hindi, Arabic, Spanish, and Chinese, to probe cross-lingual behavior. Benchmarking on GPT-5-mini reveals that it consistently favors Honesty over Autonomy across all five languages when
• 待LLM分类：否
• 标题中文：价值几何：语言模型伦理偏好对齐中的任务向量组合
• TLDR中文：大语言模型 (LLM) 越来越多地部署于必须权衡冲突道德价值观的应用中，但即便是强模型也表现出隐藏偏见和跨语言指令遵循能力的脆弱性。我们引入一个包含 12,000 个二选一困境实例的数据集，涵盖三对成对价值冲突：诚实 vs. 正义、正义 vs. 自主、自主 vs. 诚实，并提供印地语、阿拉伯语、西班牙语和中文翻译，用于探测跨语言行为。对 GPT-5-mini 的基准测试显示，在所有五种语言中它始终偏好诚实而非自主，当
• 来源文件：
• /inbox/tom/_candidates/2026-09-21-agent-rag-longcontext-candidates.json