🤖 AI Summary
Diffusion language models suffer from inefficient inference, as most tokens converge well before the final denoising step. This work proposes a training-free, token-wise early stopping mechanism that dynamically determines whether each token position has stabilized and can be frozen ahead of schedule, based on lightweight signals such as model prediction confidence and local contextual consistency. For the first time, this approach enables adaptive, per-token early termination without any fine-tuning. Evaluated across diverse benchmarks—including mathematical reasoning, general question answering, and scientific comprehension—it substantially reduces the number of diffusion steps required while preserving generation quality, achieving state-of-the-art inference efficiency.
📝 Abstract
Diffusion language models generate text through iterative refinement, a process that is often computationally inefficient because many tokens reach stability long before the final denoising step. We introduce a training-free, token-level early stopping approach that identifies convergence independently at each position. Our method leverages lightweight signals derived from the model's predictions and local context to dynamically determine when individual tokens can be finalized. This yields adaptive per-token freezing without task-specific fine-tuning, substantially reducing the total number of diffusion steps required. Across diverse benchmarks, spanning mathematical reasoning, general question answering, and scientific understanding, our approach achieves state-of-the-art efficiency gains while preserving generation quality.