Rejection Mixing: Fast Semantic Propagation of Mask Tokens for Efficient DLLM Inference

TL;DR

ReMix introduces a continuous refinement framework for diffusion large language models, significantly accelerating inference speed while maintaining quality by resolving semantic conflicts during decoding.

Contribution

It proposes ReMix, a novel, training-free method that uses continuous intermediate states and rejection rules to improve inference speed and quality in DLLMs.

Findings

Achieves 2-8x inference speedup without quality loss.

Effectively resolves semantic contradictions during decoding.

Demonstrates robustness across various experimental settings.

Abstract

Diffusion Large Language Models (DLLMs) promise fast non-autoregressive inference but suffer a severe quality-speed trade-off in parallel decoding. This stems from the ''combinatorial contradiction'' phenomenon, where parallel tokens form semantically inconsistent combinations. We address this by integrating continuous representations into the discrete decoding process, as they preserve rich inter-position dependency. We propose ReMix (Rejection Mixing), a framework that introduces a novel Continuous Mixing State as an intermediate between the initial masked state and the final decoded token state. This intermediate state allows a token's representation to be iteratively refined in a continuous space, resolving mutual conflicts with other tokens before collapsing into a final discrete sample. Furthermore, a rejection rule reverts uncertain representations from the continuous state back to the masked state for reprocessing, ensuring stability and preventing error propagation. ReMix thus mitigates combinatorial contradictions by enabling continuous-space refinement during discrete diffusion decoding. Extensive experiments demonstrate that ReMix, as a training-free method, achieves a $2-8 \times$ inference speedup without any quality degradation.