DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models

arXiv·medium signal

Presents an 8B block-diffusion language model that uses block-size curriculum learning to reliably scale parallel block-wise denoising for long chain-of-thought reasoning. Adds a concrete, named model to the growing case that diffusion LMs can compete with autoregressive models on reasoning while decoding in parallel. By Zirui Wu, Lin Zheng, Jiacheng Ye et al. (cs.CL).