Parallelism and Generation Order in Masked Diffusion Language Models: Limits Today, Potential Tomorrow

TL;DR

This paper evaluates the capabilities and limitations of Masked Diffusion Language Models (MDLMs) in parallel token generation and order flexibility, revealing current gaps and potential improvements.

Contribution

It characterizes MDLM behavior across multiple benchmarks, compares them with autoregressive models, and proposes a Generate-then-Edit paradigm to enhance their effectiveness.

Findings

MDLMs lag behind autoregressive models in dependency modeling.

Parallelism and generation order vary with task domain and reasoning stage.

MDLMs adapt their decoding order based on task requirements, such as filling easier Sudoku blanks first.

Abstract

Masked Diffusion Language Models (MDLMs) promise parallel token generation and arbitrary-order decoding, yet it remains unclear to what extent current models truly realize these capabilities. We characterize MDLM behavior along two dimensions -- parallelism strength and generation order -- using Average Finalization Parallelism (AFP) and Kendall's tau. We evaluate eight mainstream MDLMs (up to 100B parameters) on 58 benchmarks spanning knowledge, reasoning, and programming. The results show that MDLMs still lag behind comparably sized autoregressive models, mainly because parallel probabilistic modeling weakens inter-token dependencies. Meanwhile, MDLMs exhibit adaptive decoding behavior: their parallelism and generation order vary significantly with the task domain, the stage of reasoning, and whether the output is correct. On tasks that require "backward information" (e.g., Sudoku), MDLMs adopt a solution order that tends to fill easier Sudoku blanks first, highlighting their advantages. Finally, we provide theoretical motivation and design insights supporting a Generate-then-Edit paradigm, which mitigates dependency loss while retaining the efficiency of parallel decoding.