The study, led by Bing Xie and Yongjin Hu of Suzhou University together with Linlin Zhao of Suzhou No. 1 Junior High School, tackles a problem that has long frustrated reading researchers. Comprehension is not a single event but a temporal process. Readers shift between different mental operations as they move through a text: accessing word meanings, assembling sentence structure, connecting ideas across sentences, drawing inferences that go beyond what is explicitly stated, and continuously monitoring whether the whole passage still makes sense. Eye-tracking experiments and reading-time measurements offer glimpses of these operations, but conventional analytical methods tend to focus on local effects at individual words rather than estimating when one cognitive phase ends and the next begins.

The researchers’ solution is a hybrid architecture that marries two very different traditions in artificial intelligence. At its foundation sits a hierarchical hidden semi-Markov model, a probabilistic framework long used in speech recognition and bioinformatics for inferring unobserved states that persist over variable stretches of time. Unlike an ordinary hidden Markov model, which assumes the system can switch states at every single step, a semi-Markov model explicitly represents how long each state endures before a transition occurs. That property turns out to be crucial for reading, because cognitive processing does not flip at every word; a demanding inference might occupy the mind for several consecutive words, while simple word recognition flashes by almost instantly.

What makes the new model distinctive is how its parameters are generated. In classical semi-Markov models, the probabilities governing state emissions, transitions, and durations are fixed tables estimated from data. Here, the researchers let a large language model, specifically the DeBERTa-v3-base encoder, produce contextual embeddings for every reading position, drawing on the surrounding sentences to capture local word meaning, inter-sentence cohesion, and overall discourse organization. These embeddings are then mapped to the emission, transition, and duration parameters of the semi-Markov module. In other words, the probability that a reader is engaged in cross-sentence inference at a given word depends dynamically on the discourse context around that word, not on a static lookup table.

The model defines five latent cognitive states drawn from discourse-comprehension theory: lexical activation, syntactic integration, semantic connection, cross-sentence inference, and discourse monitoring. Each reading position in a text, defined as one word together with its aligned sentence and paragraph units, is assigned one dominant state. Three trained annotators labeled more than half a million reading positions across 1,200 English passages drawn from TOEFL, IELTS, and Cambridge CAE and CPE examinations, along with continuous expository and narrative texts. Annotation agreement was high, with Cohen’s kappa values ranging from 0.806 for discourse monitoring to 0.881 for lexical activation, and an overall agreement rate of 94.7 percent in the formal review sample.

During decoding, the model adds two constraints that reflect the structure of real reading. A discourse-structure consistency term encourages state boundaries to align with sentence and paragraph transitions, while a cognitive-load constraint limits erratic fluctuations in state intensity and switching frequency. The decoding process searches for the single cognitive trajectory that best balances observational fit, plausible state evolution, realistic durations, and these global constraints. The result is not just a label for each word but a complete map of the reading process: where each cognitive phase begins, how long it lasts, and how it connects to the phases before and after it.

The performance figures are striking. Across the four evaluation tasks, the model achieved a macro-F1 score of 0.860 for recognizing the dominant cognitive state at each position, a boundary accuracy of 0.83 for locating state transitions, a duration mean absolute error of just 0.52 reading units, and a correlation of 0.76 between decoded trajectories and the actual organizational structure of the passages. Ablation experiments showed that every component matters. Removing the duration term dropped macro-F1 to 0.822, and removing the discourse-coherence constraint pushed boundary accuracy down to 0.762. Crucially, sensitivity analyses confirmed that the model’s tendency to progress from low-level word processing toward high-level discourse monitoring was learned from the annotated data rather than hard-coded into the architecture.

Perhaps the most revealing findings concern the temporal fingerprints of the five states. Lexical activation proved to be the briefest state, with a mean duration of 1.74 reading units, and only 8.6 percent of its segments lasted four or more units. Discourse monitoring, by contrast, persisted for an average of 4.62 units, with 63.4 percent of its segments qualifying as long-span. Cross-sentence inference showed similarly strong persistence. This gradient mirrors what predictive-processing theories of the brain would predict: higher-level representations integrate information over broader temporal windows, while lower-level processing operates in short, rapid bursts. The model also revealed that its advantage over simpler baselines grew when texts became more complex and when the context window expanded, suggesting that deep comprehension states genuinely depend on accumulated discourse context rather than isolated sentences.

The researchers stress that the framework should not be overread as a direct model of the brain. It makes no claim of biological equivalence, does not incorporate neuronal plasticity, and lacks embodied sensorimotor interaction. Its contribution is narrower but significant: it provides a bridge between discourse-comprehension theory, probabilistic state modeling, and the contextual power of large language models. Notably, the architecture proved robust across encoder choices. When the researchers swapped DeBERTa for BERT, causal decoders from the Qwen and Gemma families, or even the state-space model Mamba, macro-F1 scores ranged from 0.818 to 0.867, indicating that the gains stem from the hybrid parameter-generation design rather than any particular language model.

Limitations remain, and the authors are candid about them. The corpus consists mainly of standardized academic reading materials, so generalization to classroom reading, open online discourse, or multimodal environments awaits further testing. The state labels come from expert annotation rather than direct measurement of readers, and the decoded trajectories have not yet been validated against eye-tracking or neural recordings. A causal variant of the model, restricted to strictly incremental left-to-right encoding like human reading, showed that high-level states suffer when future context is unavailable, quantifying how much deep comprehension relies on upcoming information. Future work, the team says, will incorporate eye-tracking and neural evidence, test reader-specific adaptation, and explore whether the framework could serve as a cognitive core for embodied language-learning agents. If those efforts succeed, the humble act of reading a paragraph may finally become a process we can watch unfold in explicit, interpretable detail.

Subject of Research: Computational modeling of cognitive states during complex English discourse reading using an LLM-parameterized hierarchical hidden semi-Markov model

Article Title: Brain-inspired cognitive state modeling of complex English discourse reading: an LLM-parameterized hierarchical semi-Markov approach

Article References: Brain-inspired cognitive state modeling of complex English discourse reading: an LLM-parameterized hierarchical semi-Markov approach. (n.d.). https://doi.org/10.3389/fnbot.2026.1930227

Image Credits: AI Generated

DOI: 10.3389/fnbot.2026.1930227

Keywords: reading comprehension, large language models, hidden semi-Markov model, cognitive modeling, discourse processing, natural language processing, predictive processing, DeBERTa, cognitive states, computational linguistics, brain-inspired AI, state duration modeling