pith. sign in

arxiv: 2604.27927 · v1 · submitted 2026-04-30 · 💻 cs.AI

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

Matteo Da Pelo , Alessio Donvito , Claudio Frongia , Pietro Salis , Antonio Lieto This is my paper

Pith reviewed 2026-05-07 07:21 UTC · model grok-4.3

classification 💻 cs.AI
keywords cognitive plausibilityAI interpretationlanguage modelsunified models of cognitionbehavioral comparisontheoretical frameworkhuman-likeness assessmentperformance analysis
0
0 comments X

The pith

A new framework shows that high AI task performance does not justify claims of human-like cognition in models like CENTAUR.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces LAPITHS to give a structured way to assess whether AI systems exhibit human-like cognition. It combines a method for scoring cognitive plausibility with direct comparisons showing that non-cognitive systems can produce similar results on the same tasks. This demonstrates that major claims for models presented as unified cognitive systems lack theoretical grounding and empirical support. A sympathetic reader would care because it separates observable behavior from inferences about internal mechanisms in AI research. If the approach is sound, many current interpretations of language model capabilities would need to be revised to avoid over-attributing mental properties.

Core claim

The authors develop LAPITHS to counteract behavioristic tendencies that equate high task performance in transformer models with human-like underlying computation. They apply two assessments: the Minimal Cognitive Grid to estimate cognitive plausibility of artificial systems, and behavioral comparisons showing that equivalent outputs can be produced by systems that do not meet structural constraints associated with cognition. This shows that claims advanced for models such as CENTAUR, proposed as artificial unified models of cognition, are not theoretically or empirically justified and that their outputs do not provide independent explanatory insight into human cognition.

What carries the argument

LAPITHS, the framework that pairs a theoretically motivated scoring of cognitive plausibility with behavioral reproduction tests using non-cognitive systems to evaluate claims of human-likeness.

If this is right

  • Cognitive claims for AI systems must include validation against structural criteria rather than relying on task performance alone.
  • Models proposed as unified accounts of cognition require explicit checks against established constraints from cognitive science.
  • Equivalent results achieved by systems without cognitive plausibility weaken inferences that the original models operate via human-like mechanisms.
  • Evaluations of transformer-based models should incorporate theoretical grounding to support assertions of cognitive abilities.
  • Research emphasis should shift toward explanatory value for human cognition instead of behavioral matching in isolation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same method could be applied to other large language models to test which capabilities truly require cognitive architectures.
  • It points toward the value of building AI systems that embed explicit cognitive constraints rather than depending on scale alone.
  • Connections emerge to broader questions about what counts as sufficient evidence for attributing computation types across artificial and natural systems.
  • Future benchmarks could integrate the plausibility criteria directly to produce more discriminating tests of model claims.

Load-bearing premise

The Minimal Cognitive Grid supplies a complete and non-circular measure of cognitive plausibility, and that reproducing similar task performance with systems lacking cognitive structural constraints is enough to invalidate claims of human-like computation.

What would settle it

An experiment in which a system that satisfies the Minimal Cognitive Grid criteria still fails to match the specific performance patterns reported for CENTAUR-like models on the same cognitive benchmarks, or in which no non-cognitive system can be constructed that matches those patterns.

Figures

Figures reproduced from arXiv: 2604.27927 by Alessio Donvito, Antonio Lieto, Claudio Frongia, Matteo Da Pelo, Pietro Salis.

Figure 1
Figure 1. Figure 1: A pictorial representation of the ascription fallacy problem, where the fact that the Humans and AI produce view at source ↗
Figure 2
Figure 2. Figure 2: Average NLL per decision for each RAG + LLM models system evaluated in this study during the two-step view at source ↗
Figure 3
Figure 3. Figure 3: Average NLL per decision for each RAG + LLM models system evaluated in this study compared with view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper introduces the LAPITHS framework to provide a theoretically grounded interpretation of AI model performances. It claims that several major assertions about models such as CENTAUR (positioned as artificial Unified Models of Cognition) lack justification, because these models fail to satisfy the structural constraints of the Minimal Cognitive Grid and because comparable task results can be reproduced by non-cognitive baseline systems whose outputs offer no independent explanatory insight into human cognition.

Significance. If the central argument holds, LAPITHS supplies a principled counter to purely behavioristic readings of transformer performance by anchoring evaluation in explicit cognitive-science constraints rather than surface metrics alone. The manuscript earns credit for defining the Minimal Cognitive Grid via explicit criteria in §3 (derived from established constraints rather than post-hoc fitting) and for illustrating the framework with concrete behavioral reproductions against non-cognitive baselines in §4; these elements make the critique falsifiable and reproducible.

minor comments (3)
  1. [Abstract] Abstract: the summary of the two quantitative assessments is clear in outline but remains high-level; adding one concrete example of a reproduced metric or one explicit Grid criterion would make the contribution more immediately graspable without lengthening the abstract.
  2. [§4] §4: the behavioral comparison protocol is described at a high level; a short table or enumerated list of the exact non-cognitive baselines, the performance metrics matched, and the statistical criterion used to declare similarity would improve reproducibility and allow readers to assess the strength of the counter-examples directly.
  3. [§3] §3: while the Grid is presented as derived from established cognitive-science constraints, a brief paragraph situating its criteria relative to prior cognitive benchmarks (e.g., those used in human cognitive modeling) would clarify its novelty and reduce any appearance of circularity for readers unfamiliar with the authors’ earlier work.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their positive and constructive review, as well as the recommendation for minor revision. The referee's summary accurately reflects the purpose of the LAPITHS framework in offering a theoretically motivated alternative to purely behavioral interpretations of high-performing AI models such as CENTAUR.

read point-by-point responses

  1. Referee: The manuscript earns credit for defining the Minimal Cognitive Grid via explicit criteria in §3 (derived from established constraints rather than post-hoc fitting) and for illustrating the framework with concrete behavioral reproductions against non-cognitive baselines in §4; these elements make the critique falsifiable and reproducible.

    Authors: We are pleased that the referee highlights these features. The criteria in §3 are drawn directly from prior cognitive-science literature on minimal requirements for cognitive plausibility, and the §4 comparisons use publicly available baselines to ensure reproducibility. We do not see a need for revision on this point. revision: no

  2. Referee: If the central argument holds, LAPITHS supplies a principled counter to purely behavioristic readings of transformer performance by anchoring evaluation in explicit cognitive-science constraints rather than surface metrics alone.

    Authors: We agree that this is the core intended contribution of the work. The framework is constructed precisely to shift evaluation away from surface performance toward structural constraints that have independent theoretical motivation in cognitive science. revision: no

Circularity Check

0 steps flagged

No significant circularity detected in LAPITHS derivation

full rationale

The paper defines LAPITHS explicitly via the Minimal Cognitive Grid in section 3, presented as derived from established cognitive science constraints rather than post-hoc fitting or self-referential inputs. Section 4 provides concrete behavioral reproductions with non-cognitive baselines that violate the grid's structural criteria while matching reported metrics. No equations or predictions reduce by construction to fitted parameters; the central claim (CENTAUR-style performance does not entail human-like computation) rests on independent application of the grid and external baselines. Any self-citations for the grid's origins are not load-bearing, as the criteria are stated explicitly and falsifiable against external benchmarks. The framework is self-contained against cognitive science literature without reducing to its own definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The central claim rests on the validity of the Minimal Cognitive Grid as an objective measure and on the assumption that matching performance with structurally different systems falsifies cognitive claims. No free parameters are described. The grid itself functions as an invented assessment entity whose independent evidence is not shown in the abstract.

axioms (2)
  • domain assumption Human cognition possesses identifiable structural constraints that can be listed in a grid and used to evaluate artificial systems.
    Invoked when the Minimal Cognitive Grid is presented as the reference point for cognitive plausibility.
  • domain assumption Reproducing task performance with systems that lack those structural constraints is sufficient to show that the original model's performance does not indicate human-like computation.
    Central to the behavioral comparison component of LAPITHS.
invented entities (2)
  • LAPITHS framework no independent evidence
    purpose: Provide a principled reference point for interpreting AI performance claims.
    Newly named method combining the grid and behavioral controls.
  • Minimal Cognitive Grid no independent evidence
    purpose: Quantitatively estimate cognitive plausibility of artificial systems.
    Core assessment tool whose construction details are not supplied in the abstract.

pith-pipeline@v0.9.0 · 5491 in / 1593 out tokens · 49096 ms · 2026-05-07T07:21:11.088790+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

9 extracted references · 7 canonical work pages · 2 internal anchors

  1. [1]

    On the Opportunities and Risks of Foundation Models

    Abbate, Francesco and Rocco Gaudenzi (2025). “Ungrounding intelligence: From biological minds to artificial cognition”. In:Philosophical Psychology, pp. 1–30. Akam, Thomas, Rui Costa, and Peter Dayan (2015). “Simple plans or sophisticated habits? State, transition and learning interactions in the two-step task”. In:PLoS computational biology11.12, e100464...

    work page internal anchor Pith review arXiv 2025

  2. [2]

    The magical number 4 in short-term memory: A reconsideration of mental storage capacity

    Springer Science & Business Media. Cowan, Nelson (2001). “The magical number 4 in short-term memory: A reconsideration of mental storage capacity”. In:Behavioral and brain sciences24.1, pp. 87–114. Crick, Francis (1989). “The recent excitement about neural networks”. In:Nature337.6203, pp. 129–132. Curtis, Clayton E and Mark D’Esposito (2003). “Persistent...

    work page doi:10.18653/v1/2025.acl-long.1415.url:https://aclanthology.org/2025.acl-long.1415/ 2001

  3. [3]

    Structure mapping in analogy and similarity

    1, pp. 59–79. 26 APREPRINT- MAY1, 2026 Gentner, Dedre and Arthur B Markman (1997). “Structure mapping in analogy and similarity.” In:American psychologist 52.1, p

    2026

  4. [4]

    doi: 10.1038/s41586-025-09422-z

    Gershman, Samuel J and Nathaniel D Daw (2017). “Reinforcement learning and episodic memory in humans and animals: an integrative framework”. In:Annual review of psychology68, pp. 101–128. Gluck, Kevin A. and John E. Laird, eds. (2018).Interactive Task Learning: Humans, Robots, and Agents Acquiring New Tasks Through Natural Interactions. Cambridge, MA: The...

    work page doi:10.1038/s41586-025-09422-z 2017

  5. [5]

    LoRA: Low-Rank Adaptation of Large Language Models

    Hick, William E (1952). “On the rate of gain of information”. In:Quarterly Journal of experimental psychology4.1, pp. 11–26. Holyoak, Keith J (2012). “Analogy and relational reasoning”. In:The Oxford handbook of thinking and reasoning, pp. 234–259. Hu, Edward J. et al. (2021).LoRA: Low-Rank Adaptation of Large Language Models. arXiv: 2106.09685 [cs.CL]. U...

    work page internal anchor Pith review arXiv 1952

  6. [6]

    Text comprehension, memory, and learning

    Kintsch, Walter (1994). “Text comprehension, memory, and learning.” In:American psychologist49.4, p

    1994

  7. [7]

    A standard model of the mind: Toward a common computational framework across artificial intelligence, cognitive science, neuroscience, and robotics

    Laird, John E, Christian Lebiere, and Paul S Rosenbloom (2017). “A standard model of the mind: Toward a common computational framework across artificial intelligence, cognitive science, neuroscience, and robotics”. In:Ai Magazine 38.4, pp. 13–26. Leivada, Evelina et al. (2025). “Large Language Model probabilities cannot distinguish between possible and im...

    work page doi:10.1093/nc/niag002 2017

  8. [8]

    David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al

    Rosenblueth, Arturo and Norbert Wiener (1945). “The role of models in science”. In:Philosophy of science12.4, pp. 316–321. Russell, Stuart Jonathan and Peter Norvig (2012).Artificial intelligence: A modern approach;[the intelligent agent book]. Prentice hall. Salis, Pietro and Matteo Da Pelo (forthcoming). “Chatting with LLMs: neither cognition nor commun...

    work page doi:10.1038/nature16961 1945

  9. [9]

    Can robots make good models of biological behaviour?

    Webb, Barbara (2001). “Can robots make good models of biological behaviour?” In:Behavioral and brain sciences 24.6, pp. 1033–1050. Wu, Charley M et al. (2018). “Generalization guides human exploration in vast decision spaces”. In:Nature human behaviour2.12, pp. 915–924. Wulff, Dirk U. and Rui Mata (2026).Addressing Longstanding Challenges in Cognitive Sci...

    work page arXiv 2001