LatentMDM Outperforms AR with KV Caching on TinyGSM

msalbergo · x · 2026-08-18

New research investigates whether masked diffusion models are truly 'any-order' models, arguing that while off-the-shelf models have predictive power, they lack genuine reasoning. The team introduces LatentMDM (latent-space masked diffusion) and the 7B-scale FlexMDM. Experiments show that LatentMDM outperforms autoregressive models with KV caching on the TinyGSM benchmark at a matched inference budget.

Original post →

More from Models

Models channel →