Dependency-Aware Revocable Decoding for Efficient Diffusion Large Language Model Inference
This work provides an incremental improvement for researchers and practitioners working with diffusion large language models, specifically addressing the trade-off between decoding parallelism and generation quality.
This paper addresses the issue of degraded generation quality in diffusion large language models (dLLMs) when increasing decoding parallelism, caused by early errors contaminating later contexts. The proposed Dependency-Aware Revocable Decoding (DARD) framework improves the speed-quality Pareto frontier, achieving a 2.71x speedup and a 4.35-point CIDEr score gain over Saber on Flickr30K.
Diffusion large language models (dLLMs) offer a promising alternative to autoregressive generation by decoding multiple tokens in parallel through iterative denoising. However, increasing decoding parallelism often degrades generation quality, as early errors can contaminate later contexts. Revocable decoding mitigates this issue by re-evaluating decoded tokens and remasking unreliable ones, but existing methods overlook that unreliable tokens may also corrupt the verification context itself. We identify this failure mode and propose Dependency-Aware Revocable Decoding (DARD), a training-free framework that separates tokens into masked, candidate, and unmasked states. DARD verifies candidate tokens using a selective context that excludes less reliable tokens and adaptively regulates their influence on subsequent decoding. Experiments across 12 textual and multimodal benchmarks on 3 open-source dLLMs show that DARD consistently improves the speed-quality Pareto frontier over recent revocable decoding methods, achieving a 2.71$\times$ speedup and a 4.35-point CIDEr score gain over Saber on Flickr30K.