Across the Levels of Analysis: Explaining Predictive Processing in Humans Requires More Than Machine-Estimated Probabilities
| Authors | Sathvik Nair & Colin Phillips |
| Year | 2026 |
| Field | NLP |
| arXiv | 2604.09466 |
| Download | |
| Categories | cs.CL |
Abstract
Under the lens of Marr's levels of analysis, we critique and extend two claims about language models (LMs) and language processing: first, that predicting upcoming linguistic information based on context is central to language processing, and second, that many advances in psycholinguistics would be impossible without large language models (LLMs). We further outline future directions that combine the strengths of LLMs with psycholinguistic models.
Engineering Breakdown
Plain English
This paper examines two fundamental claims about language models through Marr's computational levels of analysis: that predicting the next linguistic element from context is core to how language processing works, and that large language models have become essential tools for advancing psycholinguistics research. The authors critique these claims by distinguishing between what happens at different levels of analysis (computational goals, algorithmic implementation, and physical implementation), showing that prediction-as-the-goal and prediction-as-a-tool are not equivalent concepts. Rather than dismissing either perspective, they propose a research direction that combines the scalability and empirical coverage of LLMs with the mechanistic insights and cognitive grounding that psycholinguistic models provide.
Core Technical Contribution
The core contribution is a theoretical framework for evaluating claims about language models by applying Marr's three-level hierarchy of analysis to separate computational claims about language processing from implementational details about how LLMs achieve performance. This is novel because most ML research treats LLMs as black boxes for downstream tasks, whereas this paper explicitly decouples what LLMs optimize for (next-token prediction) from what human language processing actually requires (potentially richer representational goals). The authors identify a critical gap: showing that LLMs succeed at prediction doesn't prove prediction is how humans process language, nor does it invalidate psycholinguistic models that may capture human mechanisms better despite lower scale. The proposed synthesis—combining LLM capacity with psycholinguistic constraints—offers a new research direction that neither community has thoroughly explored.
How It Works
The paper applies Marr's three-level analysis as a diagnostic tool: the computational level (what problem is being solved and why), the algorithmic level (what representations and procedures implement it), and the physical/implementational level (the hardware substrate). At the computational level, the authors distinguish between prediction as the goal of human language processing versus prediction as the training objective for LLMs—these are not the same thing. They examine existing claims from psycholinguistics and LLM research, mapping each claim to the appropriate Marr level to expose which claims are being conflated. For instance, if an LLM learns a representation that correlates with brain activity during reading, that's implementational-level evidence (both use similar circuits), not proof that prediction is the computational goal. The output is a taxonomy of valid and invalid inference paths: what you can and cannot conclude when LLMs match human behavior, and where psycholinguistic models might capture constraints that LLMs bypass through scale.
Production Impact
For teams building production NLP systems, this analysis suggests that blindly scaling next-token prediction (the current LLM paradigm) may not be the right optimization target for applications requiring human-like reasoning or robustness. If you're building systems that need to match human processing constraints—dialogue systems that model turn-taking conventions, reading comprehension that respects garden-path sentences, or translation that preserves pragmatic implicature—the paper's framework shows you cannot rely on standard LLM objectives. The practical implication: consider hybrid architectures that retain LLM-scale representations but add psycholinguistic-style constraints or interpretability layers as hard requirements during training or inference. Trade-offs include increased implementation complexity, potentially higher training compute to enforce dual objectives, and latency overhead from additional constraint-checking modules. The approach is most valuable when your application domain has well-established psycholinguistic findings (reading, syntax, semantics) and when user-facing quality depends on human-like behavior, not just statistical plausibility.
Limitations and When Not to Use This
The paper does not provide concrete algorithms or trained systems demonstrating that the proposed LLM+psycholinguistics hybrid actually works better than either approach alone—it's primarily a theoretical framework without empirical validation on standard benchmarks. The authors rely on existing psycholinguistic findings, which themselves carry limitations: many are based on small populations, controlled laboratory tasks, and may not generalize to the diverse language use cases LLMs encounter at scale. Marr's framework, while elegant, is descriptive rather than prescriptive—knowing that a claim operates at the computational level doesn't immediately tell you how to build a better model. The paper also sidesteps the question of whether human language processing goals are even knowable: psycholinguistics observes behavior and brain correlates but cannot directly access the internal optimization criterion evolution implemented. Follow-up work must demonstrate that constraining LLMs with psycholinguistic priors actually improves performance on tasks where humans excel and LLMs fail, rather than merely improving interpretability.
Research Context
This work builds on decades of Marr's influence in cognitive science and extends it to a new domain—the recent explosion of LLMs—by asking whether the theoretical clarity Marr provided still applies. It sits at the intersection of neurolinguistics (how human brains process language), computational linguistics (formal models of syntax and semantics), and deep learning (empirical scaling laws for neural networks). The paper engages with classic psycholinguistic debates about garden-path sentences, incremental parsing, and predictive processing, reframing them through the lens of what can be concluded from LLM behavior. It opens a new research direction: rather than competition between psycholinguistic models (limited scale, mechanistic) and LLMs (massive scale, black-box), the field could invest in systems that inherit LLM-scale data coverage while respecting psycholinguistic constraints on representation and inference, potentially revealing which constraints are task-critical versus human-specific.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::