Diffusion language models (dLLMs) generate several tokens at once but usually throw away lots of helpful clues each step—RCD keeps and reuses those clues.
FIGR is a new way for AI to ‘think by drawing,’ using code to build clean, editable diagrams while it reasons.