Diffusion Language Model Decoding

Diffusion language models generate without a fixed left-to-right order, leaving token ordering as a central algorithmic choice. Existing systems mainly use random masking or confidence-driven ordering, which respectively suffer from train--test mismatch and myopic exploration. We introduce DPRM (Doob -transform Process Reward Model), a plug-in token-ordering module that keeps the host architecture, denoising objective and supervision unchanged, and modifies only the ordering policy. DPRM starts from confidence-driven ordering and gradually shifts to process-reward-guided ordering through online estimates. We characterize the exact DPRM policy as a reward-tilted Gibbs reveal law, prove convergence of its stagewise Soft-BoN approximation, show that the online bucketized controller tracks the exact DPRM score at empirical-Bernstein rates, and establish a sample-complexity advantage under tractable optimization assumptions. Across nine hosts covering language reasoning, test-time scaling, protein, single-cell, molecular, DNA, text-to-image generation, and VQA, DPRM order variants improve several language, DNA, and multimodal settings while also identifying boundary cases where confidence-only ordering or task-specific utilities are preferable. Code is available at: https://github.com/DakeBU/DPRM-DLLM

Discrete diffusion language models (dLLMs) enable parallel token updates with bidirectional attention, yet practical generation typically adopts blockwise semi-autoregressive decoding. This switch creates a training-inference mismatch: training denoises with full-sequence context, while inference commits tokens within a bounded block without future context. Therefore, decoding with fixed-size or heuristic-based blocks can lead to premature token commitments, as decisions are made without full access to future context that could alter those choices. Motivated by this, we propose self-containedness as a principled criterion for block commitment. A block is self-contained if its predictions remain consistent with Future-Aware (FA) or without No-Future (NF) access to future context, reframing block boundary selection as a test of self-containedness rather than a heuristic choice. Based on this principle, we introduce Variable-size Self-contained Blocks (VSB) for dLLMs. VSB scores and selects block boundaries using the divergence between token-level predictive distributions under NF and FA conditioning, which quantifies how predictions would change if future context were revealed. We provide theoretical justification linking self-containedness to predictive consistency, and extensive experiments validate VSB's efficacy over fixed-size and heuristic blockwise decoding.

While In-Context Learning (ICL) is extensively studied in Autoregressive (AR) LLMs, its mechanism within Diffusion Large Language Models (dLLMs) remains largely unexplored. Unlike AR models restricted by unidirectional causal masking, dLLMs intrinsically utilize bidirectional attention, offering extensive spatial flexibility for query placement. Unfortunately, current practices conventionally inherit AR-style trailing-query templates, often overlooking the structural paradigm shift. This paper presents a comprehensive analysis unveiling that query position is actually a first-order variable in dLLMs. Through empirical decoupling, we demonstrate that positional variance impacts generation quality on par with example semantic quality. Internally, this positional sensitivity stems from a spatial ``Recency Effect'' in attention flow and task-dependent shifts in decoding trajectories. To mitigate this instability without ground-truth labels, we reveal that traditional single-step confidence ($C_{decoded}$) fails in dLLMs. Instead, we propose Average Confidence ($\overline{C}$), a novel metric tracking the iterative decoding process. By establishing the foundational spatial ICL baselines, we introduce Auto-ICL, a training-free adaptive routing strategy that dynamically optimizes query placement, robustly approaching oracle performance across heterogeneous reasoning and perception tasks.

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 $R^{2}$-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 $R^{2}$-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.

Diffusion large language models (dLLMs) gain speed by committing multiple tokens in parallel at each denoising step, but any erroneous commitment persists as conditioning context and biases every subsequent prediction. LLaDA2.1 repairs such errors with Token-to-Token (T2T) editing, which re-examines previously unmasked tokens and overwrites them when an alternative becomes sufficiently confident. We argue that this replacement action is itself the limiting factor: under polluted context, a confident replacement can propagate the error, while under a multimodal posterior no alternative may be confident enough to trigger an edit. We propose Token-to-Mask (T2M) remasking, a training-free rule that revokes suspicious commitments by resetting them to [M] and lets the subsequent mask-filling steps re-predict them from a cleaner context. T2M improves accuracy by +13.33 points on AIME 2025 and +8.56 points on CMATH. These results suggest that, for parallel discrete generators, remasking suspect tokens rather than overwriting them is a more reliable self-correction primitive.

Discrete diffusion language models (dLLMs) have recently emerged as a promising alternative to traditional autoregressive approaches, offering the flexibility to generate tokens in arbitrary orders and the potential of parallel decoding. However, existing heuristic sampling strategies remain inefficient: they choose only a small part of tokens to sample at each step, leaving substantial room for improvement. In this work, we study the problem of token sampling order optimization and demonstrate its significant potential for acceleration. Specifically, we find that fully leveraging correct predictions at each step can reduce the number of sampling iterations by an order of magnitude without compromising accuracy. Based on this, we propose Neural Indicator Sampling (NI Sampling), a general sampling order optimization framework that utilize a neural indicator to decide which tokens should be sampled at each step. We further propose a novel trajectory-preserving objective to train the indicator. Experiments on LLaDA and Dream models across multiple benchmarks show that our method achieves up to 14.3$\times$ acceleration over full-step sampling with negligible performance drop, and consistently outperforms confidence threshold sampling in the accuracy-step trade-off. Code is available at https://github.com/imagination-research/NI-Sampling.

Diffusion large language models (dLLMs) enable parallel text generation by iteratively denoising a fully masked sequence, unmasking a subset of masked tokens at each step. Existing decoding strategies rely on static confidence metrics computed at a single denoising step, ignoring temporal history and often leading to premature unmasking of unstable tokens. In this work, we theoretically establish that a token's temporal instability, quantified by the KL divergence between consecutive prediction distributions, provides a strict lower bound on its mutual information with the remaining masked context, indicating that temporally unstable tokens are inherently unsafe to unmask. Based on this insight, we propose Stability-Weighted Decoding (SWD), a training-free, plug-and-play strategy that incorporates temporal stability into token scoring and acts as a universal modulator for arbitrary score-based decoding policies. Experiments on code generation and mathematical reasoning benchmarks demonstrate that SWD consistently improves generation accuracy across representative scoring metrics and selection policies, and exhibits exceptional robustness, maintaining a significant performance lead over standard baselines across varying acceleration ratios.

Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive language generation due to their potential for parallel decoding and global refinement of the entire sequence. To unlock this potential, DLM inference must carefully balance generation quality and decoding speed. Recent block-wise DLM decoding methods improve this trade-off by performing diffusion-based decoding sequentially in blocks. However, existing methods typically rely on fixed block schedules or current-step local signals to determine block boundaries, and use conservative confidence-based parallel decoding to avoid conflicts, limiting the quality-speed trade-off. In this paper, we argue that block-wise DLM inference requires more suitable signals for its two core decisions: cross-step signals for determining block boundaries, and token-level conflict signals for parallel decoding. Based on this view, we propose DepCap, a training-free framework for efficient block-wise DLM inference. Specifically, DepCap instantiates the cross-step signal as the influence of the last decoded block and uses it to adaptively determine how far the next block should extend, while identifying a conflict-free subset of tokens for safe parallel decoding within each block, enabling substantial inference acceleration with negligible quality degradation. DepCap is a plug-and-play method applicable to various DLMs, and compatible with existing KV-cache strategies for block-wise DLM. An information-theoretic analysis further suggests that the cumulative last-block influence on a candidate block is approximately additive across tokens, supporting the proposed block-partitioning criterion. Experimental results show that DepCap achieves favorable speed-quality trade-offs across multiple DLM backbones and reasoning and coding benchmarks, with up to 5.63$\times$ speedup without significant performance degradation.

Autoregressive vision-language models (VLMs) deliver strong multimodal capability, but their token-by-token decoding imposes a fundamental inference bottleneck. Diffusion VLMs offer a more parallel decoding paradigm, yet directly converting a pretrained autoregressive VLM into a large-block diffusion VLM (dVLM) often leads to substantial quality degradation. In this work, we present BARD, a simple and effective bridging framework that converts a pretrained autoregressive VLM into a same-architecture, decoding-efficient dVLM. Our approach combines progressive supervised block merging, which gradually enlarges the decoding block size, with stage-wise intra-dVLM distillation from a fixed small-block diffusion anchor to recover performance lost at larger blocks. We further incorporate a mixed noise scheduler to improve robustness and token revision during denoising, and memory-friendly training to enable efficient training on long multimodal sequences. A key empirical finding is that direct autoregressive-to-diffusion distillation is poorly aligned and can even hurt performance, whereas distillation within the diffusion regime is consistently effective. Experimental results show that, with $\leq$ 4.4M data, BARD-VL transfers strong multimodal capability from Qwen3-VL to a large-block dVLM. Remarkably, BARD-VL establishes a new SOTA among comparable-scale open dVLMs on our evaluation suite at both 4B and 8B scales. At the same time, BARD-VL achieves up to 3$\times$ decoding throughput speedup compared to the source model. Code is available at https://github.com/fudan-generative-vision/Bard-VL.

Diffusion large language models (DLLMs) have emerged as powerful generative models with the promise of fast text generation through parallel decoding. However, realizing this potential in practice remains challenging: reducing the number of decoding steps, typically causes a substantial degradation in output quality due to token factorization error. To alleviate this, we propose a self-distillation framework that trains a few-step student to match the generative trajectory of a full-step teacher. We theoretically and empirically show that trajectory-level supervision mitigates this factorization error, thereby enabling effective few-step decoding. We further incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that encourages mode-seeking toward the teacher's modes, yielding stronger performance on challenging reasoning tasks. Across reasoning and code-generation benchmarks, our method substantially narrows the gap between few-step and full-step decoding. The source code is available at https://github.com/Tyrion58/T3D.

Diffusion language models generate text through iterative refinement, a process that is often computationally inefficient because many tokens reach stability long before the final denoising step. We introduce a training-free, token-level early stopping approach that identifies convergence independently at each position. Our method leverages lightweight signals derived from the model's predictions and local context to dynamically determine when individual tokens can be finalized. This yields adaptive per-token freezing without task-specific fine-tuning, substantially reducing the total number of diffusion steps required. Across diverse benchmarks, spanning mathematical reasoning, general question answering, and scientific understanding, our approach achieves substantial efficiency gains while preserving generation quality.

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.

In autoregressive language models, each token is sampled by conditioning on all the past tokens; the overall string has thus been sampled from the correct underlying joint distribution represented by the model. In contrast, masked diffusion language models generate text by unmasking tokens out of order and potentially in parallel. Generating an overall string sampled from the correct underlying joint distribution would (again) require exactly one token unmasking in every full-model forward pass. The more tokens unmasked in parallel, the further away the string is from the true joint; this can be seen in the resulting drop in accuracy (but, increase in speed). In this paper we devise a way to {\em approximately} sample multiple tokens from the joint distribution in a single full-model forward pass; we do so by developing a new lightweight single-layer ``sampler" on top of an existing large diffusion LM. One forward pass of the full model can now be followed by multiple forward passes of only this sampler layer, to yield multiple unmasked tokens. Our sampler is trained to mimic exact joint sampling from the (frozen) full model. We show the effectiveness of our approximate joint sampling for both pretrained-only (Dream-7B-Base, Llada-7B-Base) and instruction-tuned (Dream-7B-Instruct, Dream-7B-Coder) models on language modeling and math & coding tasks. When four tokens are unmasked for each full-model denoising step, our sampling algorithm achieves a MAUVE score of 0.87 (vs marginal baseline of 0.31) with respect to the true joint distribution.

In this work, we propose Dimple, the first Discrete Diffusion Multimodal Large Language Model (DMLLM). We observe that training with a purely discrete diffusion approach leads to significant training instability, suboptimal performance, and severe length bias issues. To address these challenges, we design a novel training paradigm that combines an initial autoregressive phase with a subsequent diffusion phase. This approach yields the Dimple-7B model, trained on the same dataset and using a similar training pipeline as LLaVA-NEXT. Dimple-7B ultimately surpasses LLaVA-NEXT in performance by 3.9%, demonstrating that DMLLM can achieve performance comparable to that of autoregressive models. To improve inference efficiency, we propose a decoding strategy termed confident decoding, which dynamically adjusts the number of tokens generated at each step, significantly reducing the number of generation iterations. In autoregressive models, the number of forward iterations during generation equals the response length. With confident decoding, however, the number of iterations needed by Dimple is even only $\frac{\text{response length}}{3}$. We also re-implement the prefilling technique in autoregressive models and demonstrate that it does not significantly impact performance on most benchmark evaluations, while offering a speedup of 1.5x to 7x. Additionally, we explore Dimple's capability to precisely control its response using structure priors. These priors enable structured responses in a manner distinct from instruction-based or chain-of-thought prompting, and allow fine-grained control over response format and length, which is difficult to achieve in autoregressive models. Overall, this work validates the feasibility and advantages of DMLLM and enhances its inference efficiency and controllability. Code and models are available at https://github.com/yu-rp/Dimple.