A new preprint on arXiv proposes using pixel-based language models (PLMs) for reading time modeling. Unlike conventional token-based language models (TLMs), which rely on tokenization, PLMs operate directly on rendered text as pixels. The authors frame this as a practical alternative, citing flexibility as a key motivation.
The abstract outlines the general idea but does not include specific findings or comparisons in the available text. It positions PLMs as a potentially useful tool for predicting how long readers spend on text, a task where TLMs have typically been the standard approach.
Since this is a single preprint with a truncated abstract, the claims are limited to the proposal itself. No experimental results or quantitative evidence are described in the source, so the practical effectiveness of PLMs for reading time modeling remains an open question.