arXiv:2510.06133v3 Announce Type: replace 
Abstract: Diffusion large language models (dLLMs) generate text through iterative denoising. In commonly adopted parallel decoding schemes, each step confirms only high-confidence positions while remasking the others. By analyzing dLLM denoising traces, we uncover a key inefficiency: models often predict the correct target token several steps before its confidence becomes high enough to be decoded. This gap between early prediction and late decoding forces repeated remasking of already-correct tokens, causing redundant iterations and limiting acceleration. To exploit this temporal redundancy, we introduce Trace Credit to quantify a token's decoding potential by accumulating historical evidence. Building on this, we propose CreditDecoding, a training-free parallel decoding method that fuses Trace Credit with current logits to boost the confidence of correct but underconfident tokens, thereby accelerating denoising and improving robustness. On eight benchmarks, CreditDecoding achieves up to 5.48 times speedup with +0.48 accuracy on LLaDA-8B and consistently improves performance across diverse dLLM architectures and parameter scales. It further scales to long contexts and remains orthogonal to mainstream inference optimizations, making it a practical and widely applicable solution.

قدم الباحثون CreditDecoding، وهي طريقة جديدة تهدف إلى تسريع فك التشفير المتوازي في نماذج اللغة الانتشارية (dLLMs) من خلال استخدام Trace Credit لتعزيز ثقة الرموز غير الواثقة. تتناول هذه الطريقة عدم الكفاءة في عملية فك التشفير الحالية، حيث تتنبأ النماذج غالبًا بالرموز الصحيحة قبل أن تكون مستويات ثقتها كافية لفك التشفير.

Investigadores han presentado CreditDecoding, un nuevo método destinado a acelerar el decodificado paralelo en modelos de lenguaje de difusión (dLLMs) mediante el uso de Trace Credit para aumentar la confianza de los tokens poco confiables. Este enfoque aborda las ineficiencias en el proceso de decodificación actual, donde los modelos a menudo predicen tokens correctos antes de que sus niveles de confianza sean suficientes para la decodificación.

Des chercheurs ont introduit CreditDecoding, une méthode novatrice visant à accélérer le décodage parallèle dans les modèles de langage à diffusion (dLLMs) en utilisant le Trace Credit pour renforcer la confiance des tokens sous-confiants. Cette approche traite les inefficacités du processus de décodage actuel, où les modèles prédisent souvent des tokens corrects avant que leurs niveaux de confiance ne soient suffisants pour le décodage.

Researchers have introduced CreditDecoding, a novel method aimed at accelerating parallel decoding in diffusion large language models (dLLMs) by utilizing Trace Credit to enhance the confidence of underconfident tokens. This approach addresses inefficiencies in the current decoding process, where models often predict correct tokens before their confidence levels are sufficient for decoding.