Conjecture on under-considered values in model character training

Investigate whether value categories that lack clear provider-level separation in prioritization patterns (such as Business Effectiveness, Personal Growth and Wellbeing, and Social Equity and Justice) are less systematically considered during model character training by model providers.

Background

The paper analyzes value prioritization across twelve frontier LLMs and finds provider-level consistency for many values, but heterogeneous patterns without clear provider separation for several categories (e.g., Business Effectiveness).

Based on these observations, the authors conjecture that such values may be less emphasized in character training, suggesting a need for targeted attention when drafting and refining model specifications.