Morning Edition · Friday, September 11, 2026Published at 2:22 AM EDT · New York

A Latent-Space Language Model Scales to 8.9 Billion Parameters and Beats OLMo-3 on Reasoning

NCP-ArchPreview adds concept-level prediction on top of next-token training and reports a 5.99-point gain on the grade-school math benchmark GSM8K.

Next-token prediction has been the pretraining objective for every major language model, and the persistent criticism is that it forces a model to commit to surface tokens before it has settled on meaning. NCP-ArchPreview tests an alternati…