Selocus "Analyzing LLMs as Introspective Fuzzy Conceptual Agents"

Título: "Analyzing LLMs as Introspective Fuzzy Conceptual Agents: A Fibered FFCA Framework for Diagnosing Semantic Misalignment"

Ponente: Víctor Ramos González

Fecha: 19/05/2026

Hora: 11:00h

Resumen: Large Language Models (LLMs) can generate fluent and contextually plausible text while still organizing the underlying domain in ways that diverge from human-oriented ontologies.This paper addresses this form of semantic misalignment by combining guided introspective elicitation with a fibered extension of Fuzzy Formal Concept Analysis (FFCA). The proposed methodology first elicits graded object–attribute judgments from an LLM together with attenuation values associated with the model’s apparent grasp of the queried objects and attributes.It then internalizes these attenuation values by lifting objects and attributes to fibers indexed by degrees in[0,1]. The resulting fibered context supports representative sets, fibered inclusion, fibered derivations, and fibered concepts. We prove that the derivation operators form a Galois connection on representative sets and formulate a coherence criterion characterizing when an elicited attenuated context can be reconstructed as an embedding of an underlying fuzzy formal context. Studies applying the proposed methodology shows how the method can reveal category collapse, unstable hybridization, and loss of ontological discrimination in the conceptual structures projected by the model. The proposal is diagnostic rather than corrective: it does not claim direct access to internal beliefs, but provides a formal way to make introspective conceptual projections explicit, comparable, and interpretable.

Más información