💬 NLP

Timesteps of Mamba Align with Human Reading Times

This study reveals that the dynamic per-word processing timesteps of the Mamba state-space language model significantly predict human reading times, offering a novel architectural lens for understanding real-time language processing and memory retention.

Original authors: Yuji Yamamoto, Shinnosuke Isono, Yoshinobu Kawahara, Sho Yokoi

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuji Yamamoto, Shinnosuke Isono, Yoshinobu Kawahara, Sho Yokoi

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

Imagine you are reading a book. Sometimes, you breeze through a simple sentence like "The cat sat." in a flash. Other times, you hit a complex, confusing phrase and have to pause, re-read, and think hard before moving on. Your brain naturally spends more "processing time" on the hard parts.

This paper asks a fascinating question: Can a specific type of AI, called Mamba, "read" a sentence in a way that mimics how long it takes a human to read it?

Here is the breakdown of their discovery, using simple analogies.

1. The AI with a "Variable Speed" Brain

Most AI models (like the ones that power chatbots) are a bit like a metronome: they process every single word at the exact same speed, regardless of whether the word is easy or hard.

Mamba is different. It is built on a system called a "State-Space Model." Think of Mamba's brain as a smart conveyor belt that can speed up or slow down on the fly.

When the AI sees an easy word, it zooms through it quickly.
When it sees a difficult or surprising word, it slows down.

The paper focuses on a specific number inside Mamba called $\Delta t$ (Delta-t). You can think of this as the AI's internal "stopwatch." It measures how much time the AI "spends" mentally processing each word before moving to the next.

2. The Big Discovery: AI Time = Human Time

The researchers tested this stopwatch against real data. They looked at thousands of words that humans had read in experiments, recording exactly how long people paused on each word.

The Result: The AI's internal stopwatch ( $\Delta t$ ) was a surprisingly good predictor of human reading times.

When the AI's stopwatch said, "This word is tricky, I need to linger," humans usually took longer to read that word too.
When the AI said, "This is easy, let's go," humans read it quickly.

Even more impressively, this worked even when the researchers accounted for other known reasons why people pause (like word length or how predictable a word is). The AI's "time" added a new piece of the puzzle that other models missed.

3. The "Storyteller" Layers

The Mamba model has 24 layers of "neurons" (think of them as 24 different teams of workers passing notes to each other). The researchers found that not all teams were equally good at predicting human reading times.

The Local Workers: Some layers were great at handling immediate, local details (like the start of a sentence).
The Storytellers: Two specific layers (16 and 17) were the superstars. These layers seemed to hold onto the "big picture" of the story. They were the ones that best predicted how long humans took to read complex stories.
- Analogy: Imagine a team of editors. The first few editors check for typos (local details). The last few editors are the ones who remember the plot from Chapter 1 while reading Chapter 10. The paper found that these "plot-remembering" layers in the AI behaved very similarly to how human readers keep track of a story.

4. Why Does the AI Slow Down? (The Uncertainty Theory)

The paper also offers a theory on why the AI slows down. They suggest that the "stopwatch" ( $\Delta t$ ) is actually a measure of uncertainty.

The Analogy: Imagine you are walking through a foggy forest. If the path is clear, you walk fast. If the path is foggy and you aren't sure where the next step is, you slow down to look around.
The AI's Logic: The AI slows down (increases its time) when it is uncertain about what comes next or when it reaches a "boundary," like the end of a sentence. It pauses to "reset" its memory and prepare for the new context.

Interestingly, the researchers found that the AI naturally learned to slow down at the beginning of sentences, just like humans do when they start a new thought.

5. What This Means (According to the Paper)

The authors aren't saying this AI can cure reading disorders or that we should replace human teachers with it. Instead, they propose that Mamba offers a new lens for scientists.

Because Mamba has a built-in "time" variable that changes based on the input, it acts like a simulated human brain that processes information over time with a memory that is constantly updating. It allows researchers to study how "memory" and "uncertainty" interact in real-time language processing, something that was hard to see in older AI models that processed everything instantly.

In short: The paper shows that a specific AI model naturally learned to "think" at the same speed humans do, pausing longer for hard words and remembering the story's plot, giving us a new way to understand how our own brains handle language.

Technical Summary: Timesteps of Mamba Align with Human Reading Times

Problem Statement
Human language processing unfolds over time, and a central question in computational psycholinguistics is how artificial language models (LMs) represent this temporal dimension and whether their internal dynamics align with human cognitive processing. While previous studies have linked Transformer-based models to human reading times primarily through surprisal (the negative log probability of a word given context), this approach relies on the next-word prediction objective. This study investigates an alternative architecture, Mamba (a State-Space Model or SSM), which possesses an intrinsic notion of internal timesteps that dynamically modulate state transitions in response to input. The authors ask whether these input-dependent timesteps, which conceptually represent the duration of state evolution, correlate with the time human readers spend processing words.

Methodology
The study employs a multi-faceted approach combining empirical regression analysis with formal theoretical analysis of Mamba's architecture.

Datasets: The authors utilized two naturalistic reading time datasets:
- Natural Stories: Self-paced reading data from 181 native English speakers reading 10 stories.
- OneStop: Eye-tracking data from 180 native speakers reading 30 articles (both advanced and elementary versions), analyzing first-pass time and regression-path duration.
Predictors and Models:
- Primary Predictor: The discretization timestep ( $\Delta t$ ) from Mamba models (Mamba-130M and Mamba-2.8B). In Mamba, $\Delta t$ is a scalar (or vector) determined dynamically by the input token via a learned function (Eq. 5). The authors aggregated these into a word-level metric ( $\bar{\Delta}t$ ) by averaging across dimensions and taking the maximum value for subword tokens.
- Control Variables: To isolate the predictive power of $\Delta t$ , the authors controlled for low-level features (word length, corpus frequency, sentence position) and established predictors of reading time: Mamba surprisal and GPT-2 surprisal.
- Regression Analysis: Linear regression models were fitted to predict log-transformed reading times. Performance was evaluated using 10-fold cross-validation (repeated 50 times) via Mean Squared Error (MSE). The significance of $\Delta t$ was tested by comparing models with and without the variable using permutation tests, with Holm correction for multiple comparisons across layers.
Formal and Interventional Analysis:
- Linguistic Sensitivity: Correlations were computed between $\bar{\Delta}t$ and linguistic features (sentence position, syntactic distance).
- Long-Range Dependency: An intervention-based "passkey retrieval" task was used to assess which layers are essential for maintaining information over long contexts.
- Theoretical Derivation: The authors analyzed Mamba's transition matrix eigenspectrum to characterize memory retention capabilities. Furthermore, they introduced a noise term to the state transition equations to theoretically derive that the magnitude of $\Delta t$ corresponds to the conditional entropy (uncertainty) of the hidden-state transition.

Key Results

Predictive Power: The discretization timestep $\bar{\Delta}t$ is a significant predictor of human word-by-word reading times. The predictive power of the best-performing layer (e.g., Layer 17 in Mamba-130M for Natural Stories) is comparable to that of GPT-2 surprisal ( $R^2 \approx 0.22$ ).
Unique Variance: Crucially, $\bar{\Delta}t$ remains a significant predictor even when controlling for GPT-2 surprisal and other low-level variables. This indicates that Mamba's internal processing time captures variance in human reading behavior that is not explained by standard surprisal-based models.
Layer Specialization:
- Most layers show sensitivity to local linguistic features (e.g., sentence boundaries, syntactic distance).
- Layers 16 and 17 exhibit distinct behavior: they are less sensitive to local sentence structure but strongly correlate with story-level information and story position. These layers also showed the highest predictive power for Natural Stories.
- Intervention experiments confirmed that knocking out Layers 16 and 17 significantly degrades performance on long-range dependency tasks (passkey retrieval), suggesting these layers are critical for maintaining long-term context.
Theoretical Alignment:
- Memory Retention: Eigenvalue analysis of the transition matrix $exp(A)$ reveals that Layers 16 and 17 have a higher proportion of eigenvalues near 0.5 (compared to the overall distribution peaking near 0), theoretically supporting their role in long-term memory retention.
- Uncertainty: Theoretical analysis suggests that larger $\Delta t$ values correspond to higher uncertainty in state transitions. The data shows $\Delta t$ peaks at sentence boundaries, which are points of high uncertainty regarding the upcoming context, aligning with the model's need to update its state more aggressively.

Significance and Claims
The paper claims that Mamba offers a "new, valuable lens" for studying human real-time language processing. Unlike Transformers, which process sequences in parallel (or via attention mechanisms that do not inherently encode processing duration), Mamba's architecture explicitly models time through dynamic timesteps.

The authors argue that:

Alignment: The internal "processing time" ( $\Delta t$ ) of an SSM-based model aligns with human cognitive load, suggesting that the mechanism of dynamically adjusting state transition speeds is a plausible computational analogue to human reading dynamics.
Memory and Noise: The model allows researchers to investigate how different layers weigh short- vs. long-term information retention (via eigenspectrum) and how "noise" (uncertainty in state transitions) interacts with continuous memory representations.
Beyond Surprisal: The findings demonstrate that factors beyond word predictability (surprisal)—specifically, the dynamic duration of state updates—are relevant to human reading behavior.

The study concludes that while the exact learning dynamics of why $\Delta t$ peaks at specific linguistic boundaries remain an open question, the empirical and theoretical alignment between Mamba's timesteps and human reading times provides a novel framework for understanding language processing under memory constraints.