Existing evaluations of large language model (LLM) generated debates focus on argumentative quality while largely overlooking the underlying value structure. We introduce a diagnostic metric framework that compares how values are expressed and organized in human‑reference versus AI‑generated debates. The framework separates value expression—the set of recognizable values invoked—from value organization—the distribution and persistence of those values across turns. Grounded in Schwartz’s Basic‑10 value theory, debates are modeled as multi‑level value distributions over argumentative units, utterances, stances, and the whole debate, enabling metrics such as value coverage, alignment, diffusion, and cross‑turn dynamics. Across several generation systems, AI‑generated debates are broadly value‑expressive, often activating as many or more value dimensions than matched human references, but they organize these values differently. Their utterances tend to be locally value‑diffuse, and blind judges can distinguish value clarity from value focus. These findings suggest that debate evaluation should consider both value presence and value organization. Code and supplementary materials are available at https://github.com/Maoyuan-li/Value-Expression-and-Organization.
Review