New Taxonomy Classifies LLM Capabilities Across 14 Domains

A recently developed multilayer framework for assessing large language models categorizes abilities into 14 domains and 91 subskills, structured across Primitive, Constructed, and Integrative layers. This model, informed by human cognitive science rather than the architecture of LLMs, seeks to consolidate disparate evaluation findings. Researchers examined 31,505 publications from ACL, AAAI, ICML, and NeurIPS (2023–2025), annotating 15,934 papers centered on LLMs. The taxonomy indicates a significant focus on Language-Semantics, pointing out existing coverage deficiencies.