Automated Quality Assessment for LLM-Based Complex Qualitative Coding: A Confidence-Diversity Framework (2508.20462v1)

Abstract: While previous research demonstrated effective automated quality assessment for accessible LLM coding tasks, a fundamental question remains: can confidence-diversity frameworks maintain reliability for complex analytical tasks requiring specialized domain expertise and extensive text comprehension? Traditional inter-coder reliability measures become prohibitively expensive at scale, yet the lack of reliable automated quality assessment methods creates methodological barriers to AI adoption in sophisticated qualitative research. This study extends dual-signal quality assessment combining model confidence and inter-model consensus from accessible to complex analytical domains. We systematically validate this approach across three domains: legal reasoning (390 Supreme Court cases), political analysis (645 hyperpartisan articles), and medical classification (1,000 clinical transcripts). Results demonstrate that uncertainty-based indicators maintain predictive validity in complex tasks, with external entropy showing consistent negative correlations with accuracy (r = -0.179 to -0.273, p < 0.001) and confidence exhibiting positive correlations in two domains (r = 0.104 to 0.429). Systematic weight optimization achieves 6.6 to 113.7 percent improvements over single-signal approaches, with optimized weights transferring effectively across domains (100 percent success rate). An intelligent triage system reduces manual verification effort by 44.6 percent while maintaining quality standards. These findings establish that automated quality assessment can scale from accessible to complex analytical tasks, providing practical tools for expanding AI-assisted qualitative research. Future work will focus on addressing long-tail challenges in high-disagreement, low-confidence cases to further enhance screening efficiency.