Across the Levels of Analysis: Explaining Predictive Processing in Humans Requires More Than Machine-Estimated Probabilities

This paper critiques the overreliance on machine-estimated probabilities in explaining human language processing by applying Marr's levels of analysis, arguing that while predicting linguistic information is central, a complete understanding requires integrating large language models with traditional psycholinguistic frameworks.

The Gist

The Big Picture: A Debate About How We Read

Imagine you are trying to figure out how a human brain reads a sentence. For the last few years, a new tool called Large Language Models (LLMs) (like the AI you are talking to right now) has become very popular in this field.

The authors of the paper they are critiquing (Futrell & Mahowald) say: "These AI models are amazing! They predict the next word in a sentence almost perfectly. Since they do this well, they must be showing us exactly how human brains work. We should just use them to explain everything about language."

Sathvik Nair and Colin Phillips (the authors of this commentary) say: "Hold on a second. While AI is great at guessing the next word, it's not the whole story. Relying on it alone is like trying to understand a car engine just by looking at the speedometer. You miss the gears, the pistons, and the actual mechanics."

Here is the breakdown of their argument using simple analogies.

---

1. The "GPS" vs. The "Driver"

The AI's Job (The GPS):
Think of an LLM as a super-smart GPS. If you are driving and say, "I'm hungry, let's go to the...," the GPS instantly calculates the most likely destination based on millions of other people's trips. It says, "McDonald's!" with 99% confidence.

• What the AI does: It calculates probabilities. It knows what word usually comes next based on statistics.
• The Critique: The authors argue that just because the GPS knows the destination (the next word), it doesn't mean it understands how the driver (the human brain) actually steers the car, shifts gears, or reacts to a pothole.

The Human Job (The Driver):
Humans don't just guess the next word based on statistics. We have a complex, step-by-step mental process. We might pause, get confused by a tricky sentence structure, or get tricked by a linguistic illusion (like thinking a sentence makes sense when it actually doesn't).

• The Problem: The AI is a "black box." It gives the right answer, but it doesn't tell us the mechanism of how a human brain processes that answer in real-time.

2. The "Cooking Recipe" Analogy

Imagine you want to understand how a chef makes a perfect soufflé.

• The AI Approach: You look at a database of 10 million recipes. You notice that 99% of the time, the recipe says "add eggs." So, you conclude: "The secret to a soufflé is just knowing that eggs usually come next."
• The Psycholinguist Approach: You watch the chef's hands. You see them whisking, checking the temperature, and reacting when the batter falls. You realize that how they mix it matters just as much as what they mix.

The authors argue that LLMs are just the "database of recipes." They tell us what is predictable, but they fail to explain how the brain cooks the information. If we only look at the AI, we miss the "whisking" (the actual mental work).

3. The "Three Levels of Analysis" (The Ladder)

The paper uses a famous framework by a scientist named David Marr to explain why we need more than just AI. Imagine a ladder with three rungs:

1. Top Rung (The Goal): What is the system trying to do? (e.g., Predict the next word).
   • AI is great here. It tells us the goal is achieved.
2. Middle Rung (The Algorithm): How does it do it? (e.g., Does the brain use a specific memory trick? Does it get stuck on certain words?).
   • AI is weak here. It just spits out a number. It doesn't show the steps.
3. Bottom Rung (The Hardware): What is it physically made of? (e.g., Neurons firing in the brain).
   • AI is disconnected here. AI runs on silicon chips; humans run on biology.

The Authors' Point: We have been stuck on the Top Rung for too long, thinking that if we know the goal (prediction), we understand the whole system. But to truly understand human language, we need to climb down to the Middle Rung and figure out the specific "mental algorithms" that the AI is hiding.

4. The "Magic Trick" vs. The "Mechanism"

The paper mentions "linguistic illusions." Sometimes, humans read a sentence like "The key to the cabinet was rusty" and think it makes sense, even though it's grammatically weird.

• The AI might not even notice the trick because it just sees the words "key," "cabinet," and "rusty" often go together.
• The Human gets confused because their brain is actively building a structure, and that structure breaks.

If we only use AI to study language, we will never understand why humans get tricked. We need models that simulate the "breaking" of the mental structure, not just the prediction of the word.

5. The Solution: A Team Effort

The authors aren't saying "Throw away the AI." They are saying: "Don't let the AI do all the thinking."

They suggest a partnership:

• Use the AI to tell us what is statistically likely (the "what").
• Use Psycholinguistic Models (simpler, more transparent math models) to explain the mental steps (the "how").
• Use Neuroscience (brain imaging) to see the biological reality (the "hardware").

The Bottom Line

The paper is a gentle warning to the field of linguistics. It says:

"We are in love with these new AI tools because they are so good at guessing the next word. But guessing isn't the same as understanding. To truly explain how humans think and speak, we need to look under the hood, not just at the dashboard."

In short: AI is a great map, but it's not the territory. We still need to explore the terrain of the human mind ourselves.

Technical Summary

Based on the commentary by Nair and Phillips, here is a detailed technical summary of the paper "Across the Levels of Analysis: Explaining Predictive Processing in Humans Requires More Than Machine-Estimated Probabilities."

1. Problem Statement

The paper addresses a critical theoretical gap in contemporary psycholinguistics regarding the explanation of human language processing.

• The Context: There is a current resurgence of interest in computational-level accounts of language processing, driven by Large Language Models (LLMs). The prevailing view (championed by Futrell & Mahowald) suggests that human processing difficulty is primarily driven by probabilistic inference (surprisal), and that LLMs provide the necessary aggregate measures of linguistic complexity to explain this.
• The Core Issue: The authors argue that equating human processing solely with machine-estimated probabilities is a category error. While LLMs successfully model aggregate processing difficulty, they fail to capture the mechanistic (algorithmic) and biological (implementational) realities of how the human brain processes language in real-time.
• Historical Parallel: The authors draw a parallel to the 1960s Derivational Theory of Complexity (DTC). Just as DTC failed because it linked processing difficulty to syntactic transformations without understanding the underlying cognitive mechanisms, the current reliance on LLM surprisal risks repeating this mistake by focusing on "what" is predictable rather than "how" it is processed.

2. Methodology

This paper is a theoretical commentary and critical analysis rather than an empirical study with new data collection. The methodology involves:

• Marr's Levels of Analysis Framework: The authors analyze the current state of the field through David Marr's three levels:
  1. Computational Level: What is the goal of the system? (Currently dominated by LLM-based probability).
  2. Algorithmic Level: How does the system represent information and transform it? (The authors argue this is where the field is currently under-served by LLMs).
  3. Implementational Level: How is the algorithm physically realized? (Neural mechanisms).
• Literature Synthesis: The authors synthesize findings from diverse sub-fields, including:
  • Psycholinguistics: Studies on linguistic illusions, argument role reversals, and production-comprehension contrasts.
  • Neuroscience: Analysis of ERP components (N400, P600) and predictive coding theories.
  • Computational Modeling: Comparing LLMs against simpler, interpretable cognitive models (e.g., interactive activation models, predictive coding models).
• Critical Evaluation: They evaluate the limitations of LLMs in explaining specific phenomena (e.g., why LLMs conflate semantic association with contextual expectation) and contrast them with biologically plausible models.

3. Key Contributions

The paper makes several significant theoretical contributions to the debate on LLMs and psycholinguistics:

• Distinction Between Prediction and Processing: The authors clarify that while LLMs are excellent at estimating predictability (computational level), they do not explain the process of processing (algorithmic level). Human processing involves specific mechanisms (e.g., retrieval, integration, error correction) that simple probability distributions cannot capture.
• Critique of "Surprisal" as a Complete Explanation: They argue that surprisal theory, while useful for aggregate data, fails to explain:
  • Qualitative Patterns: Such as linguistic illusions (e.g., "The key to the cabinets are...") where humans make specific errors that LLMs do not replicate.
  • Argument Role Sensitivity: Humans show specific sensitivities to argument roles that LLMs often miss or handle differently.
  • Neural Dissociations: The distinct neural signatures of the N400 and P600 components suggest different cognitive processes (prediction vs. integration) that are conflated in standard LLM surprisal metrics.
• Re-elevation of the Algorithmic Level: The paper argues that the field must shift focus back to algorithmic-level questions. Progress requires characterizing the underlying mental computations (e.g., how the brain handles time delays, violated predictions, and incremental structure building) rather than just correlating difficulty with statistical probability.
• Proposal for Hybrid Models: The authors advocate for a future where LLMs are not treated as the mechanism of language processing but as components within broader, cognitively realistic models. They suggest combining LLM strengths with:
  • Predictive Coding: Models grounded in neurobiological function (e.g., Rao & Ballard, Friston) that explain neural responses more accurately than LLMs.
  • Interactivity: Models that account for interactions across different levels of representation (syntax, semantics, phonology).

4. Results and Evidence

While the paper does not present new experimental data, it synthesizes existing evidence to support its arguments:

• Limitations of LLMs:
  • LLMs fail to capture linguistic illusions (e.g., agreement attraction errors) and argument role reversals in the same way humans do.
  • LLM probabilities conflate semantic association with contextual expectations, making them poor functional explanations for specific neural responses like the N400.
  • Recent benchmarks show no evidence that LLM surprisal explains syntactic disambiguation difficulty better than simpler models.
• Success of Alternative Models:
  • Predictive Coding Models: Recent extensions of predictive coding to linguistic data (e.g., Nour Eddine et al., 2024; Ohams et al., 2026) successfully explain both controlled experimental effects and naturalistic reading data, with predictive power comparable to LLMs but with greater biological plausibility.
  • Interpretability: Simpler models (e.g., interactive activation, minimalist parsers) provide interpretable measures directly tied to cognitive processes, unlike the "black box" nature of LLMs.
• Historical Validation: The paper highlights that the field has previously moved from DTC to Surprisal by discovering finer-grained phenomena. The current reliance on LLMs risks ignoring these finer-grained insights if the focus remains solely on aggregate probability.

5. Significance

The significance of this paper lies in its corrective theoretical stance for the field of psycholinguistics and cognitive science:

• Preventing a Regression: It warns against a regression to the "DTC era" where aggregate complexity measures are mistaken for mechanistic explanations. It urges researchers not to abandon the search for algorithmic mechanisms in favor of "black box" LLM correlations.
• Redefining the Role of LLMs: It proposes a nuanced role for LLMs: they are valuable tools for estimating what is predictable (computational constraints) but must be integrated into models that explain how the brain processes that information (algorithmic and implementational levels).
• Bridging Disciplines: The paper serves as a bridge between computational linguistics and cognitive neuroscience, advocating for models that are both statistically powerful (like LLMs) and biologically plausible (like predictive coding).
• Future Direction: It outlines a roadmap for the next generation of psycholinguistic research: building process models that incorporate predictability factors but are grounded in interactive cognitive processes and neural constraints, rather than relying solely on machine-estimated probabilities.

In conclusion, Nair and Phillips argue that while LLMs have "revived" the interest in computational-level accounts, true progress in understanding human language requires moving beyond machine probabilities to understand the specific mental computations and neural mechanisms that drive human predictive processing.