CreditDecoding: Accelerating Parallel Decoding in Diffusion Large Language Models with Trace Credit

A team of researchers has introduced CreditDecoding, a training-free method that accelerates text generation in diffusion large language models by exploiting temporal redundancy in the denoising process, according to a paper published on arXiv [1]. The method, detailed in a submission last revised in May 2026, targets a specific inefficiency in how diffusion large language models (dLLMs) generate text. These models produce output through iterative denoising, where each step confirms only high-confidence tokens and remasks the rest [1]. The researchers found that models frequently predict the correct token several steps before its confidence score is high enough to be accepted, leading to repeated remasking of already-correct tokens and redundant computation [1][2]. To address this, the team developed a metric called Trace Credit, which accumulates historical evidence from the denoising process to quantify a token’s decoding potential [2]. CreditDecoding then fuses this Trace Credit with the model’s current logits to boost the confidence of correct but underconfident tokens, allowing them to be decoded earlier [1][2]. On eight benchmarks, the technique achieved up to a 5.48 times speedup with a +0.48 accuracy improvement on the LLaDA-8B model [1][2]. The authors report that the performance gains are consistent across different dLLM architectures and parameter scales, and the method scales to long-context tasks [2]. The approach is also orthogonal to other mainstream inference optimizations, meaning it can be combined with existing acceleration techniques without conflict [1][2]. The paper was submitted by Kangyu Wang and colleagues, with the initial version appearing in October 2025 and the final revision in May 2026 [1]. The work is hosted on arXiv under its Computation and Language category and is associated with experimental community projects through arXivLabs [1].