Researchers introduced LEXIC, a small recurrent neural network that estimates whether a reader understood a text by analyzing eye fixation patterns alongside word frequency and word length. The model is designed to run on-device with a compact footprint, which could enable reading interfaces that adapt to a user's comprehension in real time.
A theoretical study models sparse recurrent networks of excitatory and inhibitory units whose time constants vary across the population, using random matrix theory to analyze their dynamics. The authors report that training such networks with additive noise on working memory tasks slows and diversifies inhibitory timescales, a change linked to improved task performance.
A new preprint discusses looped neural architectures that improve problem solving by repeatedly updating an internal hidden state during inference rather than scaling parameters. The authors note that training such models is challenging because backpropagation through repeated steps is costly. The work explores how these iteration-based approaches relate to spending more compute at test time.
A new arXiv paper examines how repeating a small block of middle layers can raise a language model's effective inference depth without extra parameters or additional generated tokens. The authors propose RecurTrace, which uses adaptive latent reasoning and a loop-time memory mechanism to address two design choices in such recurrent architectures. The work is a revised preprint and reports on improving reasoning through latent recurrence.