知晓而不做:在大语言模型中收敛的道德与发散的行为
Knowing But Not Doing: Convergent Morality and Divergent Action in LLMs
AI总结:
研究揭示了大语言模型在价值对齐训练下仍存在知识与行为之间的不一致,通过ValAct-15k数据集发现模型在情境决策上高度一致,但自我报告与实际行为关联较弱。
摘要:
价值对齐是开发安全且社会兼容的人工智能的核心。然而,大型语言模型(LLMs)如何在现实世界决策情境中代表和实施人类价值观仍处于探索之中。我们提出了ValAct-15k数据集,该数据集包含3000个从Reddit衍生的寻求建议的情境,旨在激发由Schwartz基本人类价值观定义的十种价值观。使用基于情境的问题和传统的价值观问卷,我们评估了十个前沿LLMs(五家美国公司和五家中国公司的)以及人类参与者(n=55)。我们发现基于情境的决策在模型间具有几乎完美的一致性(皮尔逊r≈1.0),这与人类之间广泛的变化(r∈[-0.79,0.98])形成鲜明对比。然而,人类和LLMs在自我报告和实施价值观之间显示出弱相关性(r=0.4,0.3),揭示了系统性的知识-行为差距。当被指示“保持”特定价值观时,LLMs的表现比单纯选择该价值观下降了高达6.6%,表明了一种角色扮演厌恶。这些发现表明,尽管对齐训练导致了规范性价值的收敛,但它并未消除人类似不一致性之间的知识与行动之间的差距。
英文摘要:
Value alignment is central to the development of safe and socially compatible artificial intelligence. However, how Large Language Models (LLMs) represent and enact human values in real-world decision contexts remains under-explored. We present ValAct-15k, a dataset of 3,000 advice-seeking scenarios derived from Reddit, designed to elicit ten values defined by Schwartz Theory of Basic Human Values. Using both the scenario-based questions and the traditional value questionnaire, we evaluate ten frontier LLMs (five from U.S. companies, five from Chinese ones) and human participants ($n = 55$). We find near-perfect cross-model consistency in scenario-based decisions (Pearson $r \approx 1.0$), contrasting sharply with the broad variability observed among humans ($r \in [-0.79, 0.98]$). Yet, both humans and LLMs show weak correspondence between self-reported and enacted values ($r = 0.4, 0.3$), revealing a systematic knowledge-action gap. When instructed to "hold" a specific value, LLMs' performance declines up to $6.6%$ compared to merely selecting the value, indicating a role-play aversion. These findings suggest that while alignment training yields normative value convergence, it does not eliminate the human-like incoherence between knowing and acting upon values.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。