Published on January 24, 2026, the item describes CALM models that predict vectors for meaning blocks rather than individual tokens. It claims 4× more information per step and a 44% reduction in training compute.
On January 24, 2026, the publication described Continuous Autoregressive Language Models (CALM), which predict “next-vectors” associated with blocks of meaning rather than the next individual word piece. According to the item, each step carries 4× more information and training compute falls by 44%.
The proposal is presented as a possible alternative to token-by-token autoregressive generation. These figures and this characterization are claims made by the publication; consult the original material and check how it defines the blocks, comparison, and measured compute. If you use AI to study or apply the idea, avoid submitting personal data or confidential documents without authorization.