Ripple-Pivot Search (RPS) is a training-free decoding method for diffusion large language models (dLLMs) that identifies mid-entropy pivot positions. Proactively committing these pivots reduces uncertainty across remaining masked positions, allowing more tokens to be unmasked in parallel.
HOW THIS AFFECTS YOU
●
builderYou can potentially achieve faster inference in dLLM-based architectures using this parallel decoding method.
●
researcherThis provides a new way to exploit uncertainty reduction during the diffusion decoding process.