CreditDecoding: Accelerating Parallel Decoding in Diffusion Large Language Models with Trace Credit
Researchers have developed a new method called CreditDecoding to accelerate the text generation process in diffusion large language models (dLLMs). This technique addresses an inefficiency where models predict correct tokens earlier than their confidence scores allow for decoding, leading to redundant iterations. CreditDecoding quantifies a token's decoding potential using "Trace Credit" and fuses this with current model outputs to boost confidence in correct but underconfident tokens. This training-free approach has demonstrated significant speedups of up to 5.48 times with improved accuracy on various benchmarks and dLLM architectures. AI
IMPACT Accelerates LLM inference, potentially enabling faster and more efficient text generation for a wide range of applications.