Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper investigates model scheduling for masked diffusion language models (MDLMs) to accelerate sampling. They find that early and late denoising steps are more robust to replacement with smaller models than middle steps. By replacing the full model with a smaller one during robust steps, they achieve up to 17% FLOPs reduction with minimal impact on perplexity and sample diversity.
MDLMs can be sped up by 17% with minimal quality loss by strategically using smaller models for denoising steps that matter less.
Recent advances in masked diffusion language models (MDLMs) narrow the quality gap to autoregressive LMs, but their sampling remains expensive because generation requires many full-sequence denoising passes with a large Transformer and, unlike autoregressive decoding, cannot benefit from KV caching. In this work, we exploit the flexibility of the diffusion framework and study model scheduling, where a smaller MDLM replaces the full model at a subset of denoising steps. Across models trained on OpenWebText and LM1B, we show that early and late denoising steps are substantially more robust to such replacement than middle steps, enabling up to a 17% reduction in FLOPs with only modest degradation in generative perplexity under both unconditional and prefix-conditional generation, while preserving sample diversity. We support these findings with a step-importance analysis based on loss and KL divergence between small and large models across timesteps, as well as an exhaustive search over coarse step segments, both of which identify the middle of the diffusion trajectory as most sensitive consistently across datasets. Our results suggest that simple, architecture-agnostic scheduling rules can significantly accelerate MDLM sampling while largely preserving generation quality.