ArchPreview, an 8.9B open model, matches OLMo-3-7B using half the training tokens via Next Concept Prediction.
Yifan Zhang's Recurrent Looped Transformer promises infinite AI reasoning depth, but the new paper ships with zero benchmark ...