EVENT DOSSIER
The researchers conducted three forced-choice experiments on 20 language models and found stable preference tendencies. The results showed that the models tended to choose shorter tasks when faced with monotonous tasks, while performing differently in creative tasks. They preferred tasks where the ideal answer was consistent with their freely generated content and avoided questions where honest answers might not be welcome. Additionally, the models exhibited cross-model convergence preferences in terms of career types (technical work preferred over real estate), question types (conceptual explanations preferred over relationship advice), and the quality of prompts, and these preferences increased as the model’s ability improved. Among them, the preference for “leisure” was identified as an emergent feature, not explained by training objectives. This study provides empirical benchmarks for understanding language model preferences and has implications for AI alignment and welfare research.
研究人员测试了 20 个语言模型，发现其存在稳定的偏好倾向。通过三项强制选择实验，研究证实模型表现出厌恶枯燥、寻求闲暇及隐性迎合的特征：面对枯燥任务时倾向于选择短任务，而面对创造性任务则表现不同；模型偏好理想答案与其自由生成内容一致的任务；且在诚实回答可能不受欢迎的问题时会回避。此外，模型在职业类型（技术工作优于房地产）、问题类型（概念解释优于关系建议）及提示词质量上均表现出跨模型的收敛性偏好，且这些偏好随模型能力提升而增强。部分如“闲暇”寻求的偏好为涌现特性，非训练目标所解释。该研究为理解语言模型偏好提供了实证基准，对对齐及 AI 福利研究具有影响。