From Fixed to Flexible: Via Latent Length Awareness in dLLMs for Variable-Length Generation

The fixed-length generation paradigm of Diffusion Large Language Models (dLLMs) inherently limits their capability in tasks requiring variable-length outputs. Existing training-free approaches address this limitation through dynamic length adjustment, but often disrupt the model's learned distribution. In this work, we reveal that dLLMs possess latent length awareness: the remaining generation length can be encoded in the hidden state immediately following the given context. This latent information can be extracted from a single token, termed the Navigator Token, using a lightweight regression probe. Based on this observation, we propose NaVi, a training-free strategy that enables variable-length generation through predictive and minimally disruptive length recalibration. Specifically, NaVi initializes the generation budget via Navigator Token-based length prediction, and during decoding performs schedule-aware threshold adaptation and dynamic length recalibration with semantic prior injection to preserve distributional consistency. Experiments across diverse models and benchmarks demonstrate that NaVi improves generation quality and efficiency without task-specific tuning. Code and models will be made publicly available.