🤖 AI

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

The paper introduces the LAPITHS framework to challenge the theoretical validity of claims that AI models like CENTAUR possess human-like cognition, arguing instead that their performance stems from behavioristic patterns rather than genuine cognitive plausibility.

Original authors: Matteo Da Pelo, Alessio Donvito, Claudio Frongia, Pietro Salis, Antonio Lieto

Published 2026-05-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Matteo Da Pelo, Alessio Donvito, Claudio Frongia, Pietro Salis, Antonio Lieto

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Centaur" vs. The "Lapiths"

Imagine a new, super-smart AI named Centaur. Its creators claim it is a "Unified Model of Cognition." In plain English, they say: "We taught this AI to mimic human behavior so perfectly that it must think like a human. It's not just a calculator; it's a digital mind."

The authors of this paper, a team of researchers from Italy, say: "Hold on a minute."

They introduce a framework called LAPITHS (named after the ancient Greek warriors who fought the Centaurs). Their job is to "tame" the hype around Centaur. They argue that just because an AI acts like a human, it doesn't mean it thinks like a human. They want to separate acting (performance) from thinking (cognition).

The Core Problem: The "Ascription Fallacy"

The paper identifies a logical trap called the Ascription Fallacy.

The Analogy:
Imagine you see a robot that can perfectly mimic a human playing chess. It moves the pieces, pauses to "think," and even makes the same mistakes a human would make when tired.

The Fallacy: You look at the robot and say, "Wow, it must have a brain like a human's! It understands strategy and feels the pressure of the game."
The Reality: The robot might just be a super-fast calculator that memorized millions of chess games. It doesn't "understand" anything; it just predicts the next move based on patterns.

The authors argue that Centaur is this chess robot. It was trained on a massive dataset of human decisions (called Psych-101). Because it was trained to predict what humans would do, it predicts them very well. But the authors claim this is just behavioral mimicry, not cognitive alignment.

The Tool: The "Minimal Cognitive Grid" (MCG)

To prove their point, the authors built a scorecard called the Minimal Cognitive Grid (MCG). Think of this as a "Cognitive Truth Detector" that grades AI on three things, not just how well it answers questions.

Structure vs. Function (The "How"):
- The Question: Does the AI use the same internal "gears" as a human brain?
- The Finding: Humans learn incrementally (trial and error in real-time) and have limited working memory (we can only hold a few things in our heads). Centaur, however, was trained in batches (like reading a whole book at once) and has a massive "context window" (it can remember everything in a conversation).
- The Verdict: Centaur fails this test. It acts like a human, but its internal engine is totally different. It's like a plane flying like a bird, but using jet engines instead of flapping wings.
Generality (The "Scope"):
- The Question: Can it do many different types of thinking?
- The Finding: Centaur is good at math, reasoning, and language. But it is unimodal, meaning it only processes text. It has no eyes to see, no hands to touch, and no body to move.
- The Verdict: It gets partial points. It's smart, but it lacks the "embodied" experience of a human.
Performance Match (The "Result"):
- The Question: Does it get the right answer, make the same mistakes, and take the same amount of time?
- The Finding: Centaur is excellent at this. It predicts human choices almost perfectly and even mimics human error patterns.
- The Verdict: High score here. But the authors warn: High score here does not prove it thinks like us.

The Final Score: When you combine these, Centaur gets a low "Cognitive Plausibility" score. It's a great emulator (a perfect actor), but a poor cognitive model (a true thinker).

The Experiment: "Can Anyone Do This?"

To prove that Centaur isn't special, the researchers ran a test. They took five other powerful AI models (like GPT-4, Gemini, and others) that were NOT trained on the human data.

They gave these models a simple instruction: "Here is the rule of the game. Here is the history of what happened. Now, what would a human do?" (This is called RAG or Retrieval-Augmented Generation).

The Result:

These "untrained" models performed almost as well as Centaur.
In some cases, the difference in performance was so small it was statistically meaningless.
The Metaphor: It's like giving a group of actors a script. Centaur is the actor who memorized the script perfectly. But when you give the same script to other actors, they can also deliver the lines almost perfectly without having memorized the whole book beforehand.

Conclusion: You don't need a "special" cognitive model to mimic human behavior; you just need a smart language model and the right instructions.

The Brain Scan Test (fMRI)

Centaur's creators claimed that after training, the AI's "internal thoughts" (neural representations) started to look like human brain scans (fMRI data). They said this proved the AI was becoming "brain-like."

The authors tested this too. They asked other AI models to guess what the brain activity would look like for a specific task, just by reading the task description.

The Result:

The other models also produced guesses that correlated highly with human brain scans.
The Metaphor: If you ask a group of people to describe the feeling of "fear," they might all use similar words (heart racing, sweating). That doesn't mean they all have the exact same biological fear mechanism; it just means they are all describing the same concept.
The authors argue that high correlation in brain scans is easier to get than the creators admit. It doesn't prove the AI has a brain; it just proves it understands the concept of the task.

The Final Takeaway

The paper concludes with a warning for the AI world:

Don't confuse the map with the territory. Just because an AI draws a perfect map of human behavior (predicts what we do), it doesn't mean the AI is the territory (the human mind).
Centaur is a great predictor, but a weak explanation. It is a fantastic tool for guessing what humans will do next. But we cannot use it to explain why humans think the way they do, because its internal machinery is fundamentally different from ours.
We need better tests. We need to stop just asking "Did it get the right answer?" and start asking "Did it get the right answer using the same mental gears as a human?"

In short: Centaur is a masterful actor, but it is not a human. The authors have built a framework (LAPITHS) to ensure we don't get fooled by the performance.

1. Problem Statement

The paper addresses a critical methodological issue in contemporary AI and cognitive science: the "Ascription Fallacy." This fallacy occurs when researchers conflate human-level behavioral performance (output similarity) with human-like computational mechanisms (structural/cognitive plausibility).

The authors focus on Centaur, a recent "Unified Model of Cognition" proposed by Binz et al. (2025). Centaur is a fine-tuned Large Language Model (LLM, based on Llama 3.1 70B) trained on the Psych-101 dataset (over 10 million human decisions from 160 psychological experiments). Proponents claim Centaur not only predicts human behavior but also implements cognitively plausible mechanisms and aligns with human neural activity (fMRI).

The authors argue that Centaur's success is likely a result of functionalist optimization (matching input-output statistics) rather than structural homology (sharing underlying cognitive mechanisms). They posit that high predictive accuracy and neural correlation do not necessarily imply the system possesses human-like cognition.

2. Methodology: The LAPITHS Framework

To rigorously evaluate such claims, the authors introduce LAPITHS (Language-model Analysis through Paradigm-grounded Interpretations of Theses about Human-likenesS). This framework combines theoretical analysis with empirical replication, centered on the Minimal Cognitive Grid (MCG).

A. The Minimal Cognitive Grid (MCG)

The MCG evaluates artificial systems across three quantitative dimensions to determine "Cognitive Plausibility" ($PM$):

Functional/Structural Ratio (FSR):
- Measures the balance between functional design (output matching) and structural design (adherence to cognitive theories).
- Uses a set of structural constraints ( $c_1...c_k$ ) derived from cognitive science (e.g., working memory limits, incremental learning, model-based planning).
- Calculates a score where a low FSR indicates high structural plausibility, and a high FSR indicates the system relies on functional mimicry without structural constraints.
Generality ( $G_M$ ):
- Assesses the breadth of cognitive domains covered, based on the Cattell–Horn–Carroll (CHC) theory of cognitive abilities.
- Evaluates performance across domains like Fluid Reasoning, Quantitative Knowledge, Visual Processing, and Sensory/Motor abilities.
Performance Match ( $P_{MM}$ ):
- Compares the artificial system to human baselines on specific tasks using:
  - Accuracy: Deviation from human Negative Log-Likelihood (NLL).
  - Error Patterns: Replication of human error distributions (e.g., model-free vs. model-based strategies).
  - Execution Time: Alignment with human response times (though often unavailable for LLMs).

The final Cognitive Plausibility Score ($PM$) is a weighted sum of these dimensions ( $PM = \lambda \cdot FSR' + \mu \cdot G_M + \nu \cdot P_{MM}$ ).

B. Empirical Replication Strategy

The authors applied LAPITHS to the Two-Step Task (TST), a canonical reinforcement learning paradigm used to test model-based vs. model-free strategies.

Behavioral Replication: They tested whether standard, non-fine-tuned LLMs (using Retrieval-Augmented Generation (RAG) to access task rules) could replicate Centaur's behavioral fit (NLL) on the TST.
Neural Replication: They tested whether these same RAG-based LLMs could generate fMRI Region of Interest (ROI) activation patterns that correlate with human data, without specific neuroscientific fine-tuning.

3. Key Contributions

Theoretical Framework: Formalizes the distinction between functional adequacy and cognitive plausibility, providing a metric (MCG) to quantify the "Ascription Fallacy."
Critique of Centaur: Demonstrates that Centaur fails to satisfy key structural constraints of human cognition (specifically regarding incremental learning and memory limits), despite its high behavioral accuracy.
Empirical Challenge: Shows that "state-of-the-art" behavioral and neural alignment is not unique to specialized, fine-tuned models like Centaur but can be achieved by general-purpose LLMs via simple prompting and retrieval.

4. Key Results

A. MCG Analysis of Centaur

Structural Failure: Centaur scored very low on the Functional/Structural Ratio ( $FSR' \approx 0.18$ $F S R^{'} \approx 0.18$ ).
- It failed to implement incremental learning (it uses batch training, not online updates).
- It failed to implement model-based planning explicitly (it approximates the output via a forward pass without an internal transition model).
- It failed to respect working memory limits (its context window is unbounded, unlike human capacity).
High Performance Match: Centaur achieved a high $P_{MM}$ ($0.83$) because it was explicitly trained to match human choices.
Overall Plausibility: The composite score was low ($PM = 0.39$), indicating that while Centaur is a good predictor, it is a poor cognitive model.

B. Behavioral Replication (NLL)

RAG vs. Fine-Tuning: Several general-purpose LLMs (Llama Maverick, GPT-4o, GPT-5.1) equipped with RAG (accessing task rules) achieved NLL scores comparable to Centaur.
Statistical Significance:
- Llama Maverick (RAG): NLL $\approx 0.54$ vs. Centaur NLL $\approx 0.51$ . The difference was not statistically significant ( $p \approx 0.095$ ).
- Other models (GPT-5.1, Gemini-2.5) showed only marginal, statistically significant differences compared to Centaur, but all significantly outperformed the Llama 3.1 baseline and standard cognitive models.
Conclusion: Specialized fine-tuning is not strictly necessary to achieve human-level behavioral fit on this task.

C. Neural Alignment (fMRI)

High Correlation without Training: RAG-based LLMs (e.g., Gemini-2.5 Pro, Llama Maverick) generated ROI beta vectors that showed very high correlation with human fMRI data (Pearson $r \approx 0.93$ , Cosine Similarity $\approx 0.98$ ).
Comparison to Centaur: These values were significantly higher than the peak correlation reported for Centaur in the original study ( $r \approx 0.37$ ).
Implication: High representational alignment with human brain activity can be achieved by general models simply by processing the task structure, without needing "neural" fine-tuning. This suggests that representational similarity is a weak indicator of cognitive mechanism.

5. Significance and Conclusion

The paper concludes that the achievements of models like Centaur are over-interpreted.

Behavioral Ade $\neq$ Cognitive Plausibility: High predictive accuracy and even high neural correlation do not prove a system implements human-like cognitive mechanisms.
The Danger of Ascription: Attributing cognitive theories to AI systems based solely on output matching leads to the "Ascription Fallacy," where the AI is mistaken for a genuine model of the mind.
Future Direction: The authors advocate for a shift in AI evaluation from purely performance-driven benchmarks to theoretically grounded criteria (like the MCG) that enforce structural constraints (memory limits, learning dynamics, embodiment).
Minimal Cognitive Core: They propose the development of a "Minimal Cognitive Core" for future AI, ensuring that functional roles are realized within systems that respect the architectural constraints of biological cognition.

In essence, the paper argues that while Centaur is a powerful behavioral emulator, it is not a cognitive model, and similar performance can be achieved by less specialized systems, rendering the claim of "cognitive alignment" theoretically underdetermined.