An open-source framework for recurrent-depth language models, where a shared block of layers is applied repeatedly at inference time rather than the model using a fixed number of distinct layers. This decouples reasoning depth from parameter count, exposing test-time compute as a controllable runtime parameter rather than an architectural constant fixed at training time.
Topics: efficient transformers, low-resource NLP, test-time compute, scaling, latent reasoning.
See the paper breakdown for a plain-language walkthrough of the architecture and its trade-offs.