A Multi-Dimensional Evaluation Framework for Potemkin Behavior in Large Language Models.

Initial synthesis (Potemkin-gap observation) The initial version of the work is based on a central observation: large language models can produce correct conceptual explanations while failing to consistently apply the same concepts in structurally equivalent tasks. This phenomenon is formulated as a gap between declarative competence and applicative competence, where the ability to explain does not guarantee the ability to execute. This version remains primarily descriptive and conceptual. It highlights a structural mismatch in model behavior without proposing a formal measurement framework. The analysis is centered on the intuition of a gap and on the idea of a dissociation between apparent understanding and effective performance. 🔹 Improved version (extended multi-dimensional framework) The extended version transforms this initial observation into a formalized and reproducible evaluation framework. The gap is now quantified through a primary dimension, where declarative competence is explicitly distinguished from applicative capability. The framework introduces several complementary dimensions — memory, internal consistency, adaptation to feedback, and external verification — in order to decompose competence into independently measurable components. The goal is no longer only to observe a gap, but to structure it, measure it, and compare it across models. This version also introduces a complete evaluation architecture: a standardized dataset, a reproducible pipeline, and a multi-model benchmarking protocol. 🔹 Main difference between the two versions Initial version: qualitative observation of a behavioral gap (descriptive gap) Extended version: transformation of this gap into a multi-dimensional measurement system (quantified gap) 🔹 Summary The initial version identifies a dissociation between explanation and application. The second version formalizes this phenomenon into a structured framework that enables systematic measurement, decomposition into distinct dimensions, and comparative evaluation across large language models.