Penelope: Efficient Latent Reasoning for Transformers

Penelope introduces a new framework for pretrained decoder-only Transformers that confines recurrent computation to a designated decoder segment, facilitating efficient structured reasoning without expanding parameter size or depending on chain-of-thought token serialization. Initially, the lower decoder prefix is assessed to create a memory boundary conditioned on the problem, which is then progressively enhanced using time-modulated GRU dynamics and recurrent readout states prior to generating answers. A gradual CoT-to-latent curriculum shifts observable reasoning into this internal recurrent pathway, permitting further computation in latent space without the need to execute the entire decode repeatedly. This strategy lowers both training and deployment expenses while preserving reasoning effectiveness.