Structural Transformations in Multi-Stage Dialogues with Large Language Models – The Runport Study

This study examines structural transformations in multi-stage dialogues with large language models. It is based on a fixed five-phase protocol comprising argumentative elaboration, articulation of criteria, self-evaluation, statement categorization, and final metareflection. Only textually observable features are analyzed; no claims are made regarding internal model processes. An operational observation framework is employed to systematically capture argument structure indicators, uncertainty markers, stability criteria, and meta-references. This enables qualitative analysis that is reproducible and comparable across runs. The dataset consists of multiple complete dialogue runs. Two exemplary runs are fully documented and comparatively analyzed. Despite identical prompt architecture, differences emerge in abstraction level, localization of uncertainty, and marker construction. At the same time, the central argumentative orientation remains stable across dialogic stages. The findings show that multi-stage dialogic formats do not primarily generate substantive revision but rather functional structural shifts: arguments become more precise, uncertainties are more explicitly calibrated, and statements are systematically typologized. The contribution of this study lies not in demonstrating unexpected behavior, but in mapping dialogic transformation and identifying textual invariants under recursive evaluative conditions. The study provides a foundation for further investigations with expanded samples, varied model configurations, and combined qualitative and quantitative approaches.