Roll Out and Roll Back: Diffusion LLMs are Their Own Efficiency Teachers
Authors: Fanqin Zeng, Feng Hong, Geng Yu, Huangjie Zheng, Xiaofeng Cao, Ya Zhang, Bo Han, Yanfeng Wang, +1 more
Abstract
Diffusion Large Language Models (DLLMs) promise fast parallel generation, yet open-source DLLMs still face a severe quality-speed trade-off: accelerating decoding by revealing multiple tokens often causes substantial quality degradation. We attribute this dilemma to a train-inference mismatch amplified by irreversible decoding. While training reconstructs tokens from randomly corrupted states, efficient inference requires an adaptive denoising order, where easier tokens are revealed earlier and context-dependent ones are deferred. This view motivates two complementary methods: an inference-time method that makes parallel decoding revokable, and a training-time extension that distills the reliable order exposed by this revokable process. Accordingly, we first propose Wide-In, Narrow-Out (WINO), a training-free decoding algorithm that enables revokable parallel generation. WINO aggressively drafts multiple tokens, verifies generated tokens with enriched global context, and re-masks unreliable ones for later refinement. Building on this discovered order, we further introduce WINO+, which injects the verified denoising trajectories produced by WINO into model parameters, aligning training with efficient inference. Experiments on LLaDA and MMaDA show that WINO improves both quality and efficiency, while WINO+ further strengthens this progression. On GSM8K, WINO improves accuracy from 73.24% to 75.82% with a 6.10x step reduction, and WINO+ further achieves 76.58% with a 6.83x reduction. On Flickr30K, WINO+ reaches a 16.22x step reduction with improved CIDEr. These results demonstrate that DLLMs can serve as their own efficiency teachers by first discovering reliable denoising orders through revokable decoding and then learning to follow them for faster generation. Code is available at https://github.com/Feng-Hong/WINO-DLLM/tree/WINO-plus.
Diffusion Large Language Models (dLLMs) offer a compelling paradigm for natural language generation, leveraging parallel decoding and bidirectional attention to achieve superior global coherence compared to autoregressive models. While recent works have accelerated inference via KV cache reuse or heuristic decoding, they overlook the intrinsic inefficiencies within the block-wise diffusion process. Specifically, they suffer from spatial redundancy by modeling informative-sparse suffix regions uniformly and temporal inefficiency by applying fixed denoising schedules across all the decoding process. To address this, we propose Streaming-dLLM, a training-free framework that streamlines inference across both spatial and temporal dimensions. Spatially, we introduce attenuation guided suffix modeling to approximate the full context by pruning redundant mask tokens. Temporally, we employ a dynamic confidence aware strategy with an early exit mechanism, allowing the model to skip unnecessary iterations for converged tokens. Extensive experiments show that Streaming-dLLM achieves up to 68.2X speedup while maintaining generation quality, highlighting its effectiveness in diffusion decoding. The code is available at https://github.com/xiaoshideta/Streaming-dLLM.
Diffusion large language models (dLLMs) generate text by iteratively denoising masked token sequences. Although dLLMs can predict all masked positions in parallel within each step, the large number of denoising iterations still makes inference expensive. This cost can be reduced spatially by unmasking multiple tokens per step, or temporally by collapsing multiple denoising steps into one verification call. We propose Parallel Speculative Decoding (PSD), a training-free framework that jointly improves inference along both axes. Using the confidence scores from a single forward pass, PSD selects positions to unmask via a configurable, adaptive unmasking policy and constructs multi-depth speculative drafts without extra model calls. A final batched verification pass then applies hierarchical acceptance, keeping the deepest draft that remains consistent with the updated predictions. Experiments on three dLLMs across reasoning and code generation tasks show that PSD achieves favorable trade-offs between inference efficiency and generation quality, reaching up to 5.5× tokens per forward pass with accuracy comparable to greedy decoding.
Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to autoregressive generation by enabling parallel token prediction. However, practical dLLM decoding still suffers from high inference latency, which limits deployment. In this work, we observe that a substantial part of this inefficiency comes from recurring redundancy in the decoding process, including spatial redundancy caused by confidence clusters and positional ambiguity, and temporal redundancy caused by repeatedly remasking predictions that have already stabilized. Motivated by these patterns, we propose R2-dLLM, a unified framework for reducing decoding redundancy from both inference and training perspectives. At inference time, we introduce training-free decoding rules that aggregate local confidence and token predictions, and finalize temporally stable tokens to avoid redundant decoding steps. We further propose a redundancy-aware supervised fine-tuning pipeline that aligns the model with efficient decoding trajectories and reduces reliance on manually tuned thresholds. Experiments demonstrate that R2-dLLM consistently reduces the number of decoding steps by up to 88% compared to existing decoding strategies, while maintaining competitive generation quality across different models and tasks. These results validate that decoding redundancy is a central bottleneck in dLLMs, and that explicitly reducing it yields substantial practical efficiency gains. Our code and models are available at https://github.com/GATECH-EIC/R2-dLLM.