Paper List

期刊: ArXiv Preprint
发布日期: 2026-03-13
Human-Computer InteractionArtificial Intelligence

Developing the PsyCogMetrics™ AI Lab to Evaluate Large Language Models and Advance Cognitive Science

Marywood University | The University of Scranton | University of North Carolina Wilmington | California State University Dominguez Hills

Zhiye Jin, Yibai Li, K. D. Joshi, Xuefei (Nancy) Deng, Xiaobing (Emily) Li
Figure

30秒速读

IN SHORT: This paper addresses the critical gap between sophisticated LLM evaluation needs and the lack of accessible, scientifically rigorous platforms that integrate psychometric and cognitive science methodologies for non-technical stakeholders.

核心创新

  • Methodology Introduces the first cloud-based platform applying Classical Test Theory (CTT) and psychometric validity principles (Cronbach's α > .70, AVE > .50) to systematically evaluate LLMs as cognitive entities rather than mere tools.
  • Methodology Implements a three-cycle Action Design Science framework (Relevance-Rigor-Design) with nested Build–Intervene–Evaluate loops, bridging Popperian falsifiability, Cognitive Load Theory, and stakeholder requirements into a unified evaluation system.
  • Biology Validates that modern LLMs (GPT-4, LLaMA-3) satisfy core psychometric validity criteria—including convergent, discriminant, predictive, and external validity—and outperform earlier models (GPT-3.5, LLaMA-2) across these dimensions.

主要结论

  • The PsyCogMetrics™ AI Lab successfully operationalizes psychometric principles with demonstrated reliability metrics (Cronbach's α > .70) and validity frameworks (convergent/discriminant validity) for LLM evaluation.
  • The platform addresses three critical pain points: mitigates benchmark saturation through dynamic evaluation, reduces data contamination via reproducible workflows, and expands coverage through cognitive science methodologies.
  • Design validation shows GPT-4 and LLaMA-3 satisfy psychometric validity criteria and outperform earlier models, with GPT-4 reaching six-year-old human parity on Theory of Mind vignettes (Strachan et al., 2024).
研究空白： Current LLM evaluation suffers from benchmark saturation (new models achieve near-ceiling scores without real capability improvements), data contamination (test sets leak into training), lack of coverage for emerging capabilities, and developer-oriented tools that exclude psychology/cognitive science experts who lack programming infrastructure.

摘要: This study presents the development of the PsyCogMetrics™ AI Lab (https://psycogmetrics.ai), an integrated, cloud-based platform that operationalizes psychometric and cognitive-science methodologies for Large Language Model (LLM) evaluation. Framed as a three-cycle Action Design Science study, the Relevance Cycle identifies key limitations in current evaluation methods and unfulfilled stakeholder needs. The Rigor Cycle draws on kernel theories such as Popperian falsifiability, Classical Test Theory, and Cognitive Load Theory to derive deductive design objectives. The Design Cycle operationalizes these objectives through nested Build–Intervene–Evaluate loops. The study contributes a novel IT artifact, a validated design for LLM evaluation, benefiting research at the intersection of AI, psychology, cognitive science, and the social and behavioral sciences.