Beyond Fixed: Variable-Length Denoising for Diffusion Large Language Models

Diffusion Large Language Models (DLLMs) are emerging as a powerfulalternative to the dominant Autoregressive Large Language Models, offeringefficient parallel generation and capable global context modeling. However, thepractical application of DLLMs is hindered by a critical architecturalconstraint: the need for a statically predefined generation length. This staticlength allocation leads to a problematic trade-off: insufficient lengthscripple performance on complex tasks, while excessive lengths incur significantcomputational overhead and sometimes result in performance degradation. Whilethe inference framework is rigid, we observe that the model itself possessesinternal signals that correlate with the optimal response length for a giventask. To bridge this gap, we leverage these latent signals and introduceDAEDAL, a novel training-free denoising strategy that enables Dynamic AdaptiveLength Expansion for Diffusion Large Language Models. DAEDAL operates in twophases: 1) Before the denoising process, DAEDAL starts from a short initiallength and iteratively expands it to a coarse task-appropriate length, guidedby a sequence completion metric. 2) During the denoising process, DAEDALdynamically intervenes by pinpointing and expanding insufficient generationregions through mask token insertion, ensuring the final output is fullydeveloped. Extensive experiments on DLLMs demonstrate that DAEDAL achievesperformance comparable, and in some cases superior, to meticulously tunedfixed-length baselines, while simultaneously enhancing computational efficiencyby achieving a higher effective token ratio. By resolving the static lengthconstraint, DAEDAL unlocks new potential for DLLMs, bridging a critical gapwith their Autoregressive counterparts and paving the way for more efficientand capable generation.