Taming the Centaur(s) with LAPITHS: a framework for a theoretically grounded interpretation of AI performances

arXiv cs.AI / 5/1/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper introduces LAPITHS, a framework for theoretically grounded interpretation of AI performance, aimed at evaluating claims about “human-likeness.”
Using LAPITHS, the authors argue that several major CENTAUR-related claims (about an artificial unified model of cognition) lack theoretical and empirical support.
The work criticizes a behaviorist research tendency to treat transformer language model performance as evidence of human-like underlying computation and cognitive abilities.
LAPITHS is built around two quantitative components: the Minimal Cognitive Grid for estimating cognitive plausibility, and a behavioral comparison demonstrating that similar results can be obtained without the structural constraints tied to cognitive plausibility.
The authors conclude that some observed behaviors from CENTAUR-like systems do not independently explain human cognition and can be reproduced by other, less cognitively plausible systems.

Abstract

We introduce a framework called LAPITHS (Language model Analysis through Paradigm grounded Interpretations of Theses about Human likenesS) and use it to show that several major claims advanced by models such as CENTAUR, proposed as an artificial Unified Model of Cognition, are not theoretically or empirically justified. LAPITHS provides a principled reference point for counteracting the current behaviouristic tendency in AI research to interpret the human level performances of transformer based language models as evidence of human like underlying computation and, by extension, as signs of cognitive abilities. The novelty of LAPITHS lies in making explicit the arguments grounded in two quantitative assessments: (i) the Minimal Cognitive Grid, a theoretically motivated method for estimating the cognitive plausibility of artificial systems, and (ii) a behavioural comparison showing that results similar to those reported for CENTAUR like models can be reproduced by other systems that do not satisfy the structural constraints typically associated with cognitive plausibility, and whose outputs do not provide independent explanatory insight into human cognition.