Residual Context Diffusion (RCD) improves Diffusion Language Models by recycling computation from discarded tokens during the remasking process. This module preserves contextual information from low-confidence tokens to assist subsequent decoding iterations.
HOW THIS AFFECTS YOU
●
builderThis technique could lead to faster, more context-aware diffusion-based text generation.
●
researcherRCD presents a more efficient architecture for block-wise decoding in dLLMs.