Prior-Data Gravity and Framework Reabsorption: A Behavioral Theory of Novelty Displacement in Large Language Models

Large language models are routinely evaluated through fluency, factual accuracy, relevance, and task completion. Yet these measures do not establish whether a researcher’s conceptual framework survives generation intact. Generated text retains recognizable terminology even as the relationships, boundaries, exclusions, and organizing logic that define the framework are altered. This creates a fundamental problem for researchers who use AI systems to develop original theories and conceptual models.

This preprint investigates that problem through a naturalistic longitudinal archive of research interactions with ChatGPT, Gemini, DeepSeek, and Copilot from April 2023 to September 2026. It follows conceptual frameworks through introduction, reformulation, correction, and subsequent recurrence, examining how structural changes persist beyond the response in which a correction is accepted.

The analysis develops two central constructs: prior-data gravity, the directional pull of learned conceptual structures, and framework reabsorption, the structural reorganization of an introduced framework around familiar categories. It also examines immediate theoretical capture, epistemic direction inversion, and research obstruction, establishing framework fidelity as a distinct target for evaluating large language models.

An accompanying framework-fidelity coding grid specifies the dimensions required for systematic evaluation, while a longitudinal research agenda outlines methods for testing source preservation and correction persistence across repeated interactions. A case involving panmodal aphantasia illustrates why reproducing a framework correctly in one response does not establish its preservation throughout an interaction.

The paper is relevant to researchers in large language models, natural language processing, generative AI evaluation, machine behavior, knowledge representation, human–AI interaction, and AI-assisted scholarship. It addresses a fundamental question for AI-supported theory development: how to evaluate whether a model carries a researcher’s original conceptual contribution forward with its defining relationships intact.