The CoFI Reader: A Continuous Flow of Information approach to modeling reading

Bruno Nicenboim

Keywords: self-paced reading, Bayesian modeling, continuous flow of information, mechanistic models

Abstract:

Traditional models of reading treat comprehension as a sequence of discrete stages, in which each level of representation (for example, orthography or lexical access) is completed before the next begins. Because observed reading times are too short for fully serial processing, these models typically invoke parafoveal preview, allowing parallel processing to some extent as in E-Z Reader and SWIFT. However, such accounts focus on oculomotor control and struggle to integrate the higher-order processes that support sentence-level comprehension.

I propose the CoFI Reader (Continuous Flow of Information), which replaces strictly sequential stages with a dynamic system of concurrent processing. In CoFI, partial outputs from lower-level processes flow continuously to higher-level representations, and the highest layer modulates a stochastic timer. When the timer reaches its threshold, the reader presses a key to reveal the next word, while processing of earlier words continues to influence ongoing computations. I implement the model in a hierarchical Bayesian framework (using Stan) and fit it to self-paced reading data, a paradigm that precludes parafoveal preview.

The CoFI Reader reproduces the short latencies, the full sequence of word-level reading times, and systematic spillover effects observed in empirical studies.

Together, these results demonstrate that a mechanistic model with a continuous-flow architecture can jointly explain rapid reading times and spillover in the absence of parafoveal information, offering a principled computational alternative to stage-based models and a coherent link between cognitive processing and observable behavior.

Submission Number: 39