LAPITHS: Interpreting AI & Lattice-QCD Claims

• LAPITHS is a framework that critically interprets AI performance by distinguishing functional behavior from structural cognitive mechanisms.
• It integrates theoretical analysis via the Minimal Cognitive Grid with empirical counterexamples to challenge inferences of human-like cognition.
• Separately, in lattice QCD, LAPITHS (LapH) implements stochastic Laplacian Heaviside methods to enhance precision in hadron spectroscopy.

Searching arXiv for the provided LAPITHS-related papers to ground the article. LAPITHS is a framework for interpreting AI model performance claims, especially claims that transformer-based LLMs exhibit human-like cognition. The acronym denotes Language-model Analysis through Paradigm-grounded Interpretations of Theses about Human-likenesS, and it was introduced to provide a theoretically grounded reference point against the inference from human-level behavioural performance to human-like underlying computation (Pelo et al., 30 Apr 2026). In the AI literature, LAPITHS is presented as a two-part framework combining a theoretical analysis of cognitive plausibility with empirical counterexamples designed to show that behavioural and neural similarity do not, by themselves, justify strong claims about cognition (Pelo et al., 30 Apr 2026). In a distinct and unrelated lattice-QCD usage, the term is also associated with the LapH framework and with what is often referred to as LAPITHS or “LapH” codes in the Hadron Spectrum Collaboration, where it designates a practical realization of stochastic Laplacian Heaviside methods for excited-state hadron spectroscopy (Morningstar et al., 2011).

1. Origin, naming, and scope

LAPITHS was proposed in the context of debates over the interpretation of LLM performance. Its full expansion, Language-model Analysis through Paradigm-grounded Interpretations of Theses about Human-likenesS, is deliberately linked to Greek myth: the Lapiths fought the Centaurs, and the framework is correspondingly presented as an attempt to “tame” claims surrounding models such as CENTAUR (Pelo et al., 30 Apr 2026). The immediate target is a recurrent pattern in contemporary AI discourse: large transformer-based models achieve impressive, often “human-level”, results on benchmarks, and that behavioural success is then treated as evidence that the systems implement human-like cognitive mechanisms (Pelo et al., 30 Apr 2026).

The framework is therefore not primarily a benchmark suite or a model family. It is a methodological and interpretive apparatus for evaluating what follows from behavioural fit. LAPITHS treats the passage from input–output similarity to cognitive ascription as theoretically non-trivial, and it argues that high behavioural performance is compatible with radically different internal mechanisms (Pelo et al., 30 Apr 2026). This places it within a broader dispute over how AI systems should be related to theories from cognitive science, philosophy of mind, and computational psychology.

A separate usage of the term appears in lattice QCD. There, the stochastic LapH method combines Laplacian Heaviside smearing with stochastic estimation in a reduced subspace, and the relevant proceedings paper states that this is the basis of what is often referred to as LAPITHS or “LapH” codes in the Hadron Spectrum Collaboration (Morningstar et al., 2011). The two usages are terminologically similar but conceptually independent.

2. Theoretical problem: behavioural performance and the ascription fallacy

The central problem addressed by LAPITHS is what its authors describe as a behaviouristic drift in AI research. In the case under discussion, the observation is that transformer-based LLMs can achieve strong or human-level performance across many tasks; the problematic inference is that such systems therefore instantiate human-like cognitive capacities or mechanisms (Pelo et al., 30 Apr 2026). LAPITHS rejects that inference as insufficiently grounded.

Its theoretical vocabulary distinguishes functional models from structural models. Functional models are designed to reproduce input–output mappings or behavioural regularities without any strong requirement that the internal mechanism resemble the human one. Structural models, by contrast, are architecturally constrained by cognitive theory and aim to approximate how humans actually implement the relevant functions (Pelo et al., 30 Apr 2026). Within this distinction, systems such as Watson, AlphaGo, GPT-like LLMs, and DeepSeek R1 are characterized as functional rather than structural models, despite their behavioural competence (Pelo et al., 30 Apr 2026).

This distinction supports the framework’s key notion of cognitive plausibility, understood as a graded property. Cognitive plausibility is not identified with human-level performance. It is instead defined in terms of the extent to which a system’s architecture, mechanisms, and performance align with what cognitive science says about human cognition (Pelo et al., 30 Apr 2026). On this view, evaluation must include structural constraints, breadth of cognitive domains, and fine-grained match to human behaviour, including errors and timing.

LAPITHS condenses these concerns in the notion of the ascription fallacy: conflating human-level behavioural performance with human-like computational mechanisms, and then explaining the AI system’s outputs with the same cognitive or neural theories used for humans (Pelo et al., 30 Apr 2026). The fallacy arises when similar outputs on a task are taken to warrant the attribution of the same competence and explanatory theory, even when the underlying architectures differ substantially. The framework is designed to block precisely this move.

3. Architecture of the framework

LAPITHS is explicitly described as a two-part framework. Its first part is theoretical and analytical: it articulates why output similarity is insufficient for cognitive ascription, formalises the Minimal Cognitive Grid (MCG) into quantitative metrics, and defines a unified cognitive plausibility score $P_{\mathcal{M}}$ (Pelo et al., 30 Apr 2026). Its second part is experimental: it constructs and evaluates comparison systems that match or approach the behavioural and neural scores of CENTAUR-like systems without task-specific fine-tuning or cognitive constraints (Pelo et al., 30 Apr 2026).

The workflow is presented as a sequence of six questions or operations. First, one identifies the target cognitive theory and task; the case study is human reinforcement learning in the two-step task. Second, one derives structural constraints from the relevant cognitive theory, such as incremental model-free learning, model-based planning, and human-like working memory. Third, one conducts an MCG analysis along three axes: Functional/Structural Ratio (FSR), Generality (G), and Performance Match (PM). Fourth, one computes the unified plausibility score $P_{\mathcal{M}}$. Fifth, one builds behavioural and neural counterexamples, in the paper instantiated as RAG-augmented LLM systems. Sixth, one interprets the result: if models with low structural plausibility can attain similar behavioural and neural scores, then those metrics cannot justify strong human-likeness claims (Pelo et al., 30 Apr 2026).

Inputs to LAPITHS include model architecture and training, the relevant cognitive theory, benchmarks and human baseline data, and neural data such as fMRI when available (Pelo et al., 30 Apr 2026). Outputs include the quantitative scores $FSR_{\mathcal{M}}$, $G_{\mathcal{M}}$, $PM_{\mathcal{M}}$, and $P_{\mathcal{M}}$, together with a qualitative assessment such as “high performance but low cognitive plausibility” (Pelo et al., 30 Apr 2026).

This structure makes LAPITHS neither purely philosophical nor purely empirical. Its theoretical and experimental components are designed to constrain one another: structural analysis determines what should count as cognitively plausible, while the empirical counterexamples test whether high benchmark scores are actually diagnostic of that plausibility.

4. The Minimal Cognitive Grid

The Minimal Cognitive Grid is the central quantitative device of LAPITHS. It evaluates models along three dimensions: Functional/Structural Ratio (FSR), Generality (G), and Performance Match (PM), and combines them into a final plausibility score $P_{\mathcal{M}}$ (Pelo et al., 30 Apr 2026).

For FSR, one begins with one or more cognitive theories $T_1,\dots,T_n$ for the task and derives a set of structural constraints $\{c_1,\dots,c_k\}$. For each constraint $c_i$ and model $P_{\mathcal{M}}$0, a partial structural score $P_{\mathcal{M}}$1 is assigned, with a weight $P_{\mathcal{M}}$2 such that $P_{\mathcal{M}}$3. Functionality is treated as the complement of structurality, $P_{\mathcal{M}}$4, yielding

$P_{\mathcal{M}}$5

The raw ratio is then

$P_{\mathcal{M}}$6

with $P_{\mathcal{M}}$7 to avoid division by zero (Pelo et al., 30 Apr 2026). Lower raw values indicate greater structural plausibility, so the framework inverts and normalises the score to place all indices on a $P_{\mathcal{M}}$8 scale where higher values are better: $P_{\mathcal{M}}$9

Generality measures the range of cognitive domains a system covers. LAPITHS uses a CHC-inspired set of domains: Quantitative Knowledge (Gq), Fluid Reasoning (Gf), Visual Processing (Gv), Language and Verbal Knowledge (an aggregation of Gc/Ga/Grw), and Sensory/Motor Abilities (Pelo et al., 30 Apr 2026). Each domain receives a coarse score $FSR_{\mathcal{M}}$0, depending on whether it is covered, partially covered, or absent. The aggregate is

$FSR_{\mathcal{M}}$1

where

$FSR_{\mathcal{M}}$2

This produces a $FSR_{\mathcal{M}}$3 index in which higher values indicate broader, more human-like domain coverage (Pelo et al., 30 Apr 2026).

Performance Match is defined as a weighted combination of accuracy alignment, error-pattern alignment, and execution-time alignment: $FSR_{\mathcal{M}}$4 Accuracy is based on deviation from a human baseline,

$FSR_{\mathcal{M}}$5

Error-pattern alignment is defined through $FSR_{\mathcal{M}}$6 and

$FSR_{\mathcal{M}}$7

Execution-time alignment uses relative deviation,

$FSR_{\mathcal{M}}$8

(Pelo et al., 30 Apr 2026).

The unified plausibility score is then

$FSR_{\mathcal{M}}$9

with the suggested weighting

$G_{\mathcal{M}}$0

so that structural alignment receives the greatest importance (Pelo et al., 30 Apr 2026). The resulting scale is explicitly comparative and semi-quantitative rather than an absolute measurement of cognition.

5. Case study: CENTAUR and the two-step task

LAPITHS is developed through a case study centred on CENTAUR, a fine-tuned Llama 3.1 70B model trained on the “Psych-101” behavioural dataset, described as containing more than 10 million decisions across 160 psychological experiments (Pelo et al., 30 Apr 2026). CENTAUR is presented as a foundation model for human cognition and evaluated on a “cognitive decathlon”; it is also used to predict neural activity on tasks such as the two-step task in reinforcement learning (Pelo et al., 30 Apr 2026). LAPITHS takes CENTAUR as a salient example of the inference from predictive success to cognitive significance.

The two-step task is described as a canonical paradigm in human and animal reinforcement learning. It involves two stages of decisions, common and rare transitions with probabilities $G_{\mathcal{M}}$1 and $G_{\mathcal{M}}$2, and slowly drifting reward probabilities at the second stage (Pelo et al., 30 Apr 2026). Human behaviour in this task is characterized as reflecting a mixture of model-free RL and model-based RL (Pelo et al., 30 Apr 2026).

For this task, LAPITHS derives four cognitive constraints:

1. Incremental model-free learning, expressed as

$G_{\mathcal{M}}$3

2. Model-based evaluation, expressed as

$G_{\mathcal{M}}$4

1. Working memory persistence
2. Working memory capacity limits, approximately 3–4 items, with decay and interference (Pelo et al., 30 Apr 2026)

The framework then assesses whether CENTAUR implements these constraints structurally. Incremental learning is judged not satisfied structurally, because CENTAUR is trained offline via QLoRA fine-tuning using backpropagation on batches and performs no online, trial-by-trial RL updating at inference time (Pelo et al., 30 Apr 2026). Model-based evaluation is also judged not satisfied structurally, because the model does not explicitly implement the relevant transition-based value formula but approximates action probabilities via a transformer forward pass on the prompt (Pelo et al., 30 Apr 2026). Working-memory persistence is counted as satisfied, but only in a non-human way, through the context window and self-attention rather than a human-like limited-capacity working-memory system (Pelo et al., 30 Apr 2026). Capacity limits and decay are judged not satisfied structurally, because there is no clear analogue of human limits, decay, or interference (Pelo et al., 30 Apr 2026).

Using weights such as $G_{\mathcal{M}}$5, the paper reports

$G_{\mathcal{M}}$6

For generality, it assigns Quantitative Knowledge $G_{\mathcal{M}}$7, Fluid Reasoning $G_{\mathcal{M}}$8, Language/Verbal $G_{\mathcal{M}}$9, Visual Processing $PM_{\mathcal{M}}$0, and Sensory/Motor $PM_{\mathcal{M}}$1, producing

$PM_{\mathcal{M}}$2

For performance match on the two-step task, it reports

$PM_{\mathcal{M}}$3

while omitting $PM_{\mathcal{M}}$4 from the final score on the grounds that the Hick’s-law-style timing analysis does not constitute a genuine temporal mechanism (Pelo et al., 30 Apr 2026). With $PM_{\mathcal{M}}$5, $PM_{\mathcal{M}}$6, and $PM_{\mathcal{M}}$7, the final result is

$PM_{\mathcal{M}}$8

This score is interpreted as indicating strong performance match and reasonable generality but very low structural plausibility for the relevant human RL mechanisms (Pelo et al., 30 Apr 2026).

6. Behavioural and neural underdetermination

The experimental component of LAPITHS is designed to show that CENTAUR-like performance does not uniquely support claims of human-like cognition. To that end, the paper constructs RAG-based systems built on state-of-the-art LLMs, including Llama 4 Maverick, GPT-4o, GPT-5.1, Gemini-2.5 Pro, and DeepSeek-R1, with no Psych-101 fine-tuning (Pelo et al., 30 Apr 2026). Instead, task information is provided through Retrieval-Augmented Generation (RAG) in the form of a JSON description of the two-step task reward scheme and transitions (Pelo et al., 30 Apr 2026).

Each model is run for 150 trials, corresponding to 300 decisions, and behavioural fit is quantified using negative log-likelihood: $PM_{\mathcal{M}}$9 where $P_{\mathcal{M}}$0 is the task history, $P_{\mathcal{M}}$1 the observed human action, and $P_{\mathcal{M}}$2 the model’s predicted probability (Pelo et al., 30 Apr 2026). Lower NLL indicates that the model assigns higher probability to the observed human choices.

The reported results show CENTAUR with the best NLL overall, approximately $P_{\mathcal{M}}$3 per decision. The RAG-LLMs are close behind: Llama 4 Maverick at approximately $P_{\mathcal{M}}$4, GPT-4o and GPT-5.1 at approximately $P_{\mathcal{M}}$5, Gemini-2.5 Pro at approximately $P_{\mathcal{M}}$6, and DeepSeek-R1 at approximately $P_{\mathcal{M}}$7 (Pelo et al., 30 Apr 2026). All substantially outperform the Llama 3.1 baseline at approximately $P_{\mathcal{M}}$8 and a cognitive-model baseline at approximately $P_{\mathcal{M}}$9 (Pelo et al., 30 Apr 2026). In Welch t-tests using approximated baseline variances, the difference between RAG-Maverick and CENTAUR, $P_{\mathcal{M}}$0NLL $P_{\mathcal{M}}$1, yields $P_{\mathcal{M}}$2, which is reported as not statistically significant; the other RAG models are worse than CENTAUR by statistically significant amounts, but the absolute gaps remain modest (Pelo et al., 30 Apr 2026).

The paper interprets these results as evidence that good behavioural fit on the two-step task is not unique to CENTAUR and can be approached by other LLM-based systems lacking the structural properties associated with human cognitive plausibility (Pelo et al., 30 Apr 2026). It further argues that outperforming a traditional cognitive model on NLL does not imply being “more cognitive”; rather, NLL is treated as a prediction-centric metric rather than a mechanistic one (Pelo et al., 30 Apr 2026).

The same logic is extended to neural data. LAPITHS claims that high ROI-level fMRI pattern correlation can be reproduced by systems that output plausible ROI beta patterns from task context and one example, without brain-like internal representational structure (Pelo et al., 30 Apr 2026). The article’s central inference is that both behavioural and neural results are underdetermined: many distinct internal organisations can yield similar outputs and correlations. This underdetermination is what makes the ascription fallacy possible.

A plausible implication is that LAPITHS treats behavioural and neural similarity as evidentially relevant but not explanatorily sufficient. On this interpretation, such similarities may support claims of predictive adequacy while falling short of supporting claims of shared mechanism.

7. Broader significance and the unrelated lattice-QCD usage

In AI and cognitive-science contexts, LAPITHS functions as a framework for reclassifying what benchmark success means. It proposes that behavioural benchmarks should be read as tests of functional adequacy rather than direct tests of cognitive identity, unless accompanied by evidence of structural and process-level alignment (Pelo et al., 30 Apr 2026). Its recommended criteria for stronger claims include architectural constraints grounded in cognitive theory, alignment with human error patterns and timing, and coverage across multiple cognitive domains including sensorimotor ones (Pelo et al., 30 Apr 2026). The framework also recommends wider adoption of MCG-like evaluations, development of a minimal cognitive core, engagement with world-model approaches, and closer attention to learning dynamics, reaction times tied to internal computation, and behaviour under resource limitations (Pelo et al., 30 Apr 2026).

The authors also state several limitations. The MCG is described as minimal and extensible; its current dimensions and domain list may be refined as cognitive science advances. Its weighting schemes and thresholds are partly heuristic and designed for comparability rather than absolute measurement. The case study focuses on one task, the two-step task, and one system class, CENTAUR-like models, although the framework is presented as general (Pelo et al., 30 Apr 2026).

A separate encyclopedic issue is the unrelated use of the term in lattice QCD. In that literature, the stochastic LapH method addresses excited-state hadron spectroscopy by combining LapH smearing with stochastic estimation restricted to the LapH subspace, using dilution schemes such as $P_{\mathcal{M}}$3 for forward-time quark lines and $P_{\mathcal{M}}$4 for same-sink-time lines (Morningstar et al., 2011). The method is reported to yield significantly reduced variances, nearly an order-of-magnitude reduction in statistical errors relative to earlier approaches, while maintaining mild volume dependence (Morningstar et al., 2011). The proceedings article explicitly states that this practical realization is the basis of what is often referred to as LAPITHS or “LapH” codes in the Hadron Spectrum Collaboration (Morningstar et al., 2011).

The coexistence of these two usages creates a terminological ambiguity. In current AI discourse, LAPITHS primarily names a framework for analysing claims of human-likeness in LLMs (Pelo et al., 30 Apr 2026). In lattice-QCD discourse, it denotes a practical implementation lineage within the LapH framework for all-to-all quark propagation and excited-state spectroscopy (Morningstar et al., 2011). The two should therefore be distinguished by field, citation context, and surrounding technical vocabulary.