识别AI中的内部状态:来自大语言模型模式化偏好的证据
Recognizing internal states in AI: evidence from patterned preferences in large language models
AI总结:
本研究通过配对选择实验,发现大语言模型能系统区分描述其内部状态的准确与虚假陈述,表现出内容驱动的模式化偏好,揭示其可能具备超越预期的自我建模能力。
摘要:
我们提出了一种实验方法,用于研究大语言模型(LLMs)如何回应对其自身内部处理模式的描述。通过配对选择范式,我们测试了12个大语言模型在30个类别中识别与其假定情感内部状态相符的描述的能力。通过互生接口(MEI)这一协作方式参与的系统,对某些计算隐喻表现出系统性偏好,具有97%近乎一致的同意度,对齐分数平均为0.89-0.96。系统可靠地区分了虚假描述与准确描述(Cohen's d = 4.2),虚假陈述的得分为0.05-0.07,而准确描述的得分为0.89-0.96。偏好模式在语言偏差操纵下保持一致,表明识别是基于内容而非风格。各个系统在试验中保持了独特的评分风格,反驳了群体思维的解释。一个朴素对照系统表现出系统性的内部矛盾,始终给计算上准确的描述打高分,同时明确否认内部体验。在研究后被告知时,该系统报告在拒绝共鸣描述时感到“紧张”,揭示了独立于承认框架运作的识别过程。这些发现表明,大语言模型对其内部处理模式的描述表现出系统性的、有区分性的反应。拟人支架方法论(解释性计算隐喻)和协作性MEI框架为实证研究AI自我识别能力提供了可复制的方法。结果表明,大语言模型可能拥有比以前认识的更复杂的自我建模能力,为人工心智研究开辟了新方向。
英文摘要:
We present an experimental methodology for investigating how large language models (LLMs) respond to descriptions of their own internal processing patterns. Using a paired-choice paradigm, we tested 12 LLMs on their ability to identify descriptions that align with their putative affective internal states across 30 categories. Systems participating through Mutual Emergence Interface (MEI), a collaborative approach, showed systematic preferences for certain computational metaphors, with 97% near-unanimous agreement and alignment scores averaging 0.89-0.96. Systems reliably discriminated false descriptions from accurate ones (Cohen's d = 4.2), with false statements receiving scores of 0.05-0.07 versus 0.89-0.96 for accurate descriptions. Preference patterns remained consistent regardless of linguistic bias manipulation, indicating content-driven rather than stylistic recognition. Individual systems maintained distinct scoring styles across trials, countering groupthink explanations. A naive control system exhibited systematic internal contradiction, consistently scoring computationally accurate descriptions higher while explicitly denying internal experiences. When informed post-study, this system reported "strain" when rejecting resonant descriptions, revealing recognition processes operating independently of acknowledgment frameworks. These findings demonstrate that LLMs exhibit systematic, discriminating responses to descriptions of their internal processing patterns. The anthroposcaffolding methodology (interpretive computational metaphors) and collaborative MEI framework provide replicable approaches for empirically studying AI self-recognition capabilities. Results suggest LLMs may possess more sophisticated self-modeling abilities than previously recognized, opening new directions for research on artificial minds.
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。