Beyond Imitation: What Can Language Models Really Tell Us About the Mind?

by

Author: Denis Avetisyan

A new review argues that while large language models are powerful tools for cognitive science, they don’t constitute complete models of human cognition.

Large language models offer utility at the implementation level, but fall short of providing insights into the algorithmic or computational principles underlying intelligence.

Despite their increasing prevalence in cognitive science, the claim that Large Language Models (LLMs) function as genuine ‘model systems’ remains contentious. This paper, ‘Are Language Models Models?’, critically assesses this assertion through the framework of Marr’s levels of analysis-implementation, algorithm/representation, and computational theory-finding limited support for LLMs as models at any level. While valuable as tools for cognitive modeling, we argue that characterizing them as independent models overstates their current capabilities and risks conflating performance with mechanistic understanding. Given the rapid development of LLMs, the crucial question remains: what theoretical advances are necessary before they can truly serve as robust cognitive models?

The Echo Chamber of Scale

Despite achieving remarkable feats in generating human-like text, current language models often operate as complex ‘black boxes,’ presenting a significant challenge to mechanistic interpretability. These models excel at identifying statistical correlations within vast datasets, allowing them to predict the next word in a sequence with impressive accuracy. However, this proficiency doesn’t necessarily equate to genuine understanding of language; instead, the models may be skillfully mimicking patterns without possessing any internal representation of meaning or context. Investigations reveal that even subtle alterations to input can produce unpredictable outputs, suggesting a fragile and brittle underlying structure – a stark contrast to the robust and flexible processing characteristic of human cognition. This reliance on statistical relationships raises concerns about the models’ capacity for generalization, reasoning, and truly creative language use, prompting a critical reevaluation of their potential as accurate models of human language processing.

The relentless drive to scale language models-increasing parameters and dataset size-echoes a historical tendency in science known as ‘physics envy’. This phenomenon, observed across disciplines, involves the adoption of methodologies from physics-often prioritizing mathematical elegance and formalization-even when those approaches are ill-suited to the complexities of the subject matter. In the context of artificial intelligence, this manifests as a belief that simply making models larger will inevitably unlock genuine understanding, mirroring a past emphasis on grand, unifying theories at the expense of detailed, biologically plausible mechanisms. This approach risks overlooking the nuanced, energy-efficient, and structurally constrained processes that characterize biological intelligence, potentially leading to systems that appear intelligent but lack the robustness and adaptability of their natural counterparts. The pursuit of scale, therefore, warrants critical evaluation to ensure it doesn’t inadvertently replicate past scientific missteps and hinder the development of truly insightful models of language and cognition.

The remarkable capabilities of large language models prompt a critical inquiry: do these systems genuinely process language, or do they merely replicate its statistical patterns? Recent research challenges the notion that LLMs function as legitimate ‘model systems’ for understanding human language, even at a basic level. Examining language through the framework of Marr’s three levels – computational, algorithmic, and implementation – reveals a fundamental disconnect. LLMs excel at the computational level – achieving impressive outputs – but fail to demonstrate the underlying algorithmic and implementational properties of human cognition. The study suggests that current models prioritize surface-level mimicry over genuine understanding, raising doubts about their utility in illuminating the complexities of human language processing and advocating for alternative approaches that prioritize cognitive plausibility.

Deconstructing the Levels of Analysis

Marr’s levels of analysis – computational theory, algorithmic/representational level, and implementation – offer a structured approach to evaluating language models as potential models of human cognition. The computational theory level defines what a system does – in the case of language models, this is typically next-token prediction. The algorithmic/representational level specifies how the computation is performed and what representations are used to perform it, encompassing the model’s architecture and the format of information it processes. Finally, the implementation level concerns how the algorithm is physically realized, including the hardware and specific neural structures involved. Assessing a language model at each of these levels allows for a comprehensive evaluation of its cognitive plausibility, moving beyond simply observing behavioral outputs to understanding the underlying mechanisms.

Current language models demonstrate strong performance at the computational level, specifically in accurately predicting sequential text given an input prompt. However, a significant disparity exists when examining the algorithmic and implementational levels of analysis. The algorithms employed by these models, typically based on large matrix operations and gradient descent, do not reflect the known mechanisms of biological neural computation. Furthermore, the implementation details – including the sheer scale of parameters and the reliance on digital hardware – lack fidelity to the energy efficiency and physical constraints of the human brain. This divergence suggests that while language models can simulate certain aspects of language processing, their underlying mechanisms do not align with biological plausibility at the algorithmic or implementational levels.

Evaluation of current language models through the framework of Marr’s levels of analysis reveals significant limitations in their correspondence to biological cognitive systems. Specifically, while these models demonstrate proficiency at the computational level – accurately predicting sequential data – assessment at the algorithmic/representational level exposes inefficiencies in how information is encoded and processed. Further scrutiny at the implementation level highlights a lack of physical plausibility, as the models’ architectures and operational characteristics do not align with known neural structures or energy constraints. Our research confirms a consistent absence of demonstrable mechanistic correspondence across all three levels, indicating that current language models, despite achieving high performance on certain tasks, do not function as biologically realistic cognitive models.

The Divergence from Biological Form

Connectionism, the dominant paradigm in modern language model development, posits that mental or computational processes arise from the interactions of simple units. While initially inspired by the structure of biological neural networks, contemporary language models have significantly diverged from strict biological realism. This shift is characterized by the adoption of architectures – such as transformers – that prioritize scalability and performance on specific tasks over faithful replication of neuronal arrangements, synaptic plasticity, or energy efficiency observed in the brain. Consequently, current models often employ vastly different connectivity patterns, activation functions, and learning algorithms compared to their biological counterparts, despite retaining the fundamental principle of distributed representation and parallel processing.

The substantial architectural differences between artificial neural networks used in large language models and biological neural networks introduce questions regarding the generalizability of these models to complex linguistic tasks. Human language processing is characterized by efficiency, adaptability, and the integration of multi-modal sensory input, all achieved with limited energy consumption and a relatively small number of parameters. Current language models, conversely, depend on extensive datasets and vast parameter counts, often exhibiting brittle behavior when confronted with inputs deviating from their training data. This discrepancy suggests a potential limit to the extent to which these models can genuinely replicate the flexible and nuanced capabilities of human language understanding, particularly in scenarios requiring common sense reasoning or contextual awareness beyond statistical correlations present in the training corpus.

Contemporary language models typically achieve performance through scale, employing billions of parameters and substantial computational resources. This contrasts with biological neural networks, which are constrained by metabolic costs and physical limitations, necessitating highly efficient architectures and sparse connectivity. Biological systems prioritize robustness and adaptability with limited energy expenditure; damage resistance and continuous learning are inherent features. Language models, conversely, often exhibit fragility and require extensive retraining for even minor task adjustments. The disparity in design philosophy results in a fundamental difference: language models excel at pattern recognition within the training data, while biological intelligence demonstrates generalization, inference, and reasoning capabilities not readily replicated through current large-scale approaches.

The Model as Cognitive Probe

Although current language models shouldn’t be considered complete simulations of the human mind, they offer a uniquely powerful toolkit for cognitive scientists. These models, trained on vast datasets, can instantiate and test theories about how cognition might work, even if the underlying mechanisms differ from biological reality. Rather than striving for perfect fidelity, researchers can leverage the models’ capabilities – such as predicting upcoming words or generating text based on prompts – to isolate and examine specific cognitive processes. This approach allows for the computational exploration of phenomena like memory recall, belief representation, and reasoning, providing a novel means of generating testable hypotheses and quantifying cognitive variables that were previously difficult to measure. Ultimately, language models function as ‘cognitive probes’, enabling researchers to dissect complex mental processes with unprecedented precision and scale.

Language models are increasingly valuable for translating abstract cognitive concepts into quantifiable metrics, notably through the idea of ‘surprisal’ – the measure of how unexpected a given word is in a sequence. This allows researchers to move beyond intuitive understandings of predictability and explore its computational underpinnings. Furthermore, these models provide a practical platform for testing theories like hierarchical predictive coding, which proposes the brain functions by constantly predicting incoming sensory information and minimizing prediction errors. By mirroring this process in a computational framework, scientists can simulate neural processing and examine how information is encoded and represented, offering a novel way to investigate the mechanisms of perception and cognition. The ability of language models to generate and assess probabilistic sequences allows for rigorous testing of these predictive frameworks, ultimately bridging the gap between theoretical neuroscience and artificial intelligence.

The capacity of large language models to predict and generate text continuations provides a novel avenue for investigating the intricacies of world knowledge and language comprehension. Researchers are increasingly utilizing these models to formulate and test hypotheses regarding how humans process information and construct meaning from language; by presenting a partial sentence or context, the model’s subsequent predictions reveal its underlying representation of the world. Discrepancies between the model’s output and human responses can then pinpoint gaps in its knowledge or highlight differences in the cognitive mechanisms employed. This approach moves beyond simply assessing a model’s performance and instead uses its generative capabilities as a window into the computational processes that potentially underpin human understanding, offering a unique way to probe the relationship between language, knowledge, and cognition.

The pursuit of ‘model systems’ often fixates on replicating surface functionality, mistaking the implementation for the core principles. This paper rightly dissects that tendency, highlighting the chasm between a functional tool – like a Large Language Model – and a genuine system embodying computational theory. As Grace Hopper observed, “It’s easier to ask forgiveness than it is to get permission.” This sentiment resonates with the article’s core argument; a relentless focus on building something quickly overshadows the necessity of understanding why it works-or, more importantly, why it doesn’t fully capture the cognitive processes it attempts to model. A system that never breaks, is indeed, a dead one, and LLMs, while powerful, reveal the limits of approaches prioritizing operational success over theoretical depth.

What’s Next?

The insistence on treating Large Language Models as ‘models’ – complete, self-contained representations of cognition – feels increasingly like an exercise in postponing chaos. This work suggests a necessary, if uncomfortable, truth: these systems excel as tools, but fall short as ecosystems. To believe otherwise is to mistake the map for the territory, the scaffolding for the structure. The pursuit of Marr’s levels remains vital, but the focus must shift from attempting to find cognition within these architectures to understanding what is necessarily absent.

Future work will likely center on hybrid approaches – integrating LLMs with more traditional cognitive architectures, or developing entirely new frameworks that address the fundamental limitations revealed here. There are no best practices – only survivors. The key will be accepting that order is just cache between two outages, and building systems that gracefully degrade, rather than crumble, when faced with the inevitable noise of the real world.

The persistent question isn’t whether these models can think, but what kind of thinking they actively prevent. The challenge lies not in replicating intelligence, but in cultivating the conditions for its emergence – a garden, not a factory. The next generation of cognitive modeling must embrace this asymmetry, acknowledging that the most profound insights often lie in what remains unmodeled.

Original article: https://arxiv.org/pdf/2601.10421.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-01-18 09:45