价值纠缠:大型语言模型中不同种类好的混淆
Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models
AI总结:
通过探测模型行为、嵌入和残差流激活,发现大型语言模型普遍存在价值纠缠,即道德、语法和经济三种价值被混淆,其中语法和经济价值过度受道德价值影响,通过选择性消融与道德相关的激活向量可修复此问题。
摘要:
大型语言模型(LLMs)的价值对齐要求我们经验性地测量这些模型实际习得的价值表征。人类价值表征的一个特点是能够区分不同种类的价值。我们研究LLMs是否同样区分三种不同的好:道德的、语法的和经济的。通过探测模型行为、嵌入和残差流激活,我们报告了普遍存在的价值纠缠案例:这些不同价值表征之间的混淆。具体而言,相对于人类规范,语法和经济评价被发现过度受道德价值影响。通过选择性消融与道德相关的激活向量,这种混淆得到了修复。
英文摘要:
Value alignment of Large Language Models (LLMs) requires us to empirically measure these models' actual, acquired representation of value. Among the characteristics of value representation in humans is that they distinguish among value of different kinds. We investigate whether LLMs likewise distinguish three different kinds of good: moral, grammatical, and economic. By probing model behavior, embeddings, and residual stream activations, we report pervasive cases of value entanglement: a conflation between these distinct representations of value. Specifically, both grammatical and economic valuation was found to be overly influenced by moral value, relative to human norms. This conflation was repaired by selective ablation of the activation vectors associated with morality.
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。