Just on Time: Token-Level Early Stopping for Diffusion Language Models

TL;DR

This paper presents a training-free, token-level early stopping method for diffusion language models that adaptively finalizes tokens during generation, significantly improving efficiency without sacrificing quality.

Contribution

It introduces a novel, task-agnostic early stopping technique that dynamically determines token convergence at each step, reducing computation in diffusion models.

Findings

Achieves state-of-the-art efficiency gains across multiple benchmarks.

Maintains high generation quality despite reduced diffusion steps.

Applicable without task-specific fine-tuning.

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.