Researchers have introduced STDec, a novel decoding approach for Diffusion Large Language Models (dLLMs) that enhances efficiency and maintains generation quality. Unlike previous methods that rely on global confidence thresholds, STDec incorporates spatial and temporal stability to guide the decoding process. This involves dynamically adjusting token-adaptive thresholds based on neighboring decoded states and the consistency of predicted token IDs across denoising steps. STDec is training-free and compatible with existing acceleration techniques, demonstrating significant speedups, such as a 14.17x improvement on the MBPP benchmark with LLaDA, while preserving generation scores. AI
IMPACT This new decoding method could lead to faster and more efficient inference for Diffusion Large Language Models.
RANK_REASON The cluster contains an academic paper detailing a new method for dLLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
