Diffusion Language Model Decoding

We present DMax, a new paradigm for efficient diffusion language models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On two H200 GPUs, our model achieves an average of 1,338 TPS at batch size 1. Code is available at: https://github.com/czg1225/DMax

Vision-language models (VLMs) predominantly rely on autoregressive decoding, which generates tokens one at a time and fundamentally limits inference throughput. This limitation is especially acute in physical AI scenarios such as robotics and autonomous driving, where VLMs are deployed on edge devices at batch size one, making AR decoding memory-bandwidth-bound and leaving hardware parallelism underutilized. While block-wise discrete diffusion has shown promise for parallel text generation, extending it to VLMs remains challenging due to the need to jointly handle continuous visual representations and discrete text tokens while preserving pretrained multimodal capabilities. We present Fast-dVLM, a block-diffusion-based VLM that enables KV-cache-compatible parallel decoding and speculative block decoding for inference acceleration. We systematically compare two AR-to-diffusion conversion strategies: a two-stage approach that first adapts the LLM backbone with text-only diffusion fine-tuning before multimodal training, and a direct approach that converts the full AR VLM in one stage. Under comparable training budgets, direct conversion proves substantially more efficient by leveraging the already multimodally aligned VLM; we therefore adopt it as our recommended recipe. We introduce a suite of multimodal diffusion adaptations, block size annealing, causal context attention, auto-truncation masking, and vision efficient concatenation, that collectively enable effective block diffusion in the VLM setting. Extensive experiments across 11 multimodal benchmarks show Fast-dVLM matches its autoregressive counterpart in generation quality. With SGLang integration and FP8 quantization, Fast-dVLM achieves over 6x end-to-end inference speedup over the AR baseline.

Diffusion Large Language Models (dLLMs) have achieved rapid progress, viewed as a promising alternative to the autoregressive paradigm. However, most dLLM decoders still adopt a global confidence threshold, and do not explicitly model local context from neighboring decoded states or temporal consistency of predicted token IDs across steps. To address this issue, we propose a simple spatio-temporal stability guided decoding approach, named STDec. We observe strong spatio-temporal stability in dLLM decoding: newly decoded tokens tend to lie near decoded neighbors, and their predicted IDs often remain consistent across several denoising steps. Inspired by this stability, our STDec includes spatial-aware decoding and temporal-aware decoding. The spatial-aware decoding dynamically generates the token-adaptive threshold by aggregating the decoded states of nearby tokens. The temporal-aware decoding relaxes the decoding thresholds for tokens whose predicted token IDs remain consistent over denoising steps. Our STDec is training-free and remains compatible with cache-based acceleration methods. Across textual reasoning and multimodal understanding benchmarks, STDec substantially improves throughput while maintaining comparable task performance score. Notably, on MBPP with LLaDA, STDec achieves up to 14.17x speedup with a comparable score. Homepage: https://yzchen02.github.io/STDec.

Test-time scaling investigates whether a fixed diffusion language model (DLM) can generate better outputs when given more inference compute, without additional training. However, naive best-of-$K$ sampling is fundamentally limited because it repeatedly draws from the same base diffusion distribution, whose high-probability regions are often misaligned with high-quality outputs. We propose $S^3$ (Stratified Scaling Search), a classical verifier-guided search method that improves generation by reallocating compute during the denoising process rather than only at the final output stage. At each denoising step, $S^3$ expands multiple candidate trajectories, evaluates them with a lightweight reference-free verifier, and selectively resamples promising candidates while preserving diversity within the search frontier. This procedure effectively approximates a reward-tilted sampling distribution that favors higher-quality outputs while remaining anchored to the model prior. Experiments with LLaDA-8B-Instruct on MATH-500, GSM8K, ARC-Challenge, and TruthfulQA demonstrate that $S^3$ consistently improves performance across benchmarks, achieving the largest gains on mathematical reasoning tasks while leaving the underlying model and decoding schedule unchanged. These results show that classical search over denoising trajectories provides a practical mechanism for test-time scaling in DLMs.

Discrete masked diffusion language models such as LLaDA generate text through iterative denoising, where mask tokens are progressively replaced with predicted tokens. LLaDA2.1 introduced a Token-to-Token (T2T) editing mechanism that accelerates generation by directly replacing committed tokens suspected of being incorrect. However, we identify fundamental limitations of T2T editing: it couples error detection with replacement, pollutes the generation context with potentially incorrect tokens, and introduces a train-inference noise mismatch where systematic model-generated errors differ from the random perturbations seen during training. We propose Token-to-Mask (T2M) remasking, a training-free, drop-in replacement for T2T editing that resets suspected erroneous tokens back to the mask state, allowing the diffusion process to re-predict them under cleaner context. We design and empirically validate three complementary error detection strategies -- probability-based, trigger-mirrored, and temporal-difference-based -- and provide a unified theoretical analysis showing that T2M remasking purifies the generation context, converts systematic inference errors back to the model's native mask noise type, and enables delayed commitment for joint multi-position optimization. Comprehensive experiments across 12 benchmarks spanning knowledge, reasoning, mathematics, coding, and instruction following show that T2M generally improves performance on tasks requiring precise token-level output, with the largest gain on mathematics (+5.92% on CMATH). Error analysis on CMATH reveals that the dominant failure mode is last-mile token corruption -- where correct reasoning produces a corrupted final answer -- and that T2M repairs 59.4% of such cases.

Diffusion language models (DLMs) are promising alternatives to autoregressive language models (ARMs), yet the intrinsic differences in their generated text remain underexplored. We first find empirically that off-the-shelf DLMs exhibit lower $n$-gram entropy, higher semantic coherence, and higher semantic diversity. To understand the cause, we conduct controlled experiments that decouple the effects of training objectives and decoding algorithms. Results suggest that the DLM training objective contributes to the increases in semantic coherence and semantic diversity, but has a minor influence on entropy. These differences are primarily driven by the bidirectional context; other components in the training objective, such as input masking, label masking, and the weighting function, have a much weaker influence. Further, our experiments demonstrate that the reduction in entropy stems from DLMs' decoding algorithms, particularly confidence-based remasking strategies. We provide a theoretical understanding for this entropy reduction phenomenon. Together, our work uncovers key mechanisms underlying the differences between DLMs and ARMs in text generation, and informs future design of training objectives and decoding algorithms in DLMs.

Discrete diffusion language models (dLLMs) accelerate text generation by unmasking multiple tokens in parallel. However, parallel decoding introduces a distributional mismatch: it approximates the joint conditional using a fully factorized product of per-token marginals, which degrades output quality when selected tokens are strongly dependent. We propose DEMASK (DEpendency-guided unMASKing), a lightweight dependency predictor that attaches to the final hidden states of a dLLM. In a single forward pass, it estimates pairwise conditional influences between masked positions. Using these predictions, a greedy selection algorithm identifies positions with bounded cumulative dependency for simultaneous unmasking. Under a sub-additivity assumption, we prove this bounds the total variation distance between our parallel sampling and the model's joint. Empirically, DEMASK achieves 1.7-2.2$\times$ speedup on Dream-7B while matching or improving accuracy compared to confidence-based and KL-based baselines.

Diffusion language models (DLMs) offer a promising path toward low-latency generation through parallel decoding, but their practical efficiency depends heavily on the decoding trajectory. In practice, this advantage often fails to fully materialize because standard training does not provide explicit supervision over token reveal order, creating a train-inference mismatch that leads to suboptimal decoding behavior. We propose Trajectory-Ranked Instruction Masked Supervision (TRIMS), a simple trajectory-guided supervised fine-tuning framework that injects trajectory supervision into standard Masked Diffusion Language Model (MDLM) training with minimal overhead. Instead of relying on costly DLM-based distillation, TRIMS uses lightweight signals from an autoregressive teacher to guide a trajectory-aware masking strategy, encouraging the model to learn more effective decoding orders. Experiments on LLaDA and Dream across math and coding benchmarks show that TRIMS significantly improves the accuracy-parallelism trade-off over both standard MDLM training and train-free acceleration baselines, while achieving competitive performance with prior distillation-based approaches at substantially lower training cost. Further analysis shows that TRIMS leads to better decoding trajectories, validating the effectiveness of trajectory-guided supervision for DLMs.

Diffusion large language models (dLLMs) theoretically permit token decoding in arbitrary order, a flexibility that could enable richer exploration of reasoning paths than autoregressive (AR) LLMs. In practice, however, random-order decoding often hurts generation quality. To mitigate this, low-confidence remasking improves single-sample quality (e.g., Pass@$1$) by prioritizing confident tokens, but it also suppresses exploration and limits multi-sample gains (e.g., Pass@$k$), creating a fundamental quality--exploration dilemma. In this paper, we provide a unified explanation of this dilemma. We show that low-confidence remasking improves a myopic proxy for quality while provably constraining the entropy of the induced sequence distribution. To overcome this limitation, we characterize the optimal distribution that explicitly balances quality and exploration, and develop a simple Independent Metropolis--Hastings sampler that approximately targets this distribution during decoding. Experiments across a range of reasoning benchmarks including MATH500, AIME24/25, HumanEval, and MBPP show that our approach yields better exploration-quality tradeoff than both random and low-confidence remasking.

Block-diffusion language models offer a promising path toward faster-than-autoregressive generation by combining block-wise autoregressive decoding with within-block parallel denoising. However, in the few-step regime needed for practical acceleration, standard confidence-thresholded decoding is often brittle: aggressive thresholds hurt quality, while conservative thresholds require unnecessary denoising steps. Existing approaches that address this issue either require additional training or incur extra test-time compute. We present S2D2, a training-free self-speculative decoding framework for block-diffusion language models. Our key observation is that a block-diffusion model becomes autoregressive when the block size is reduced to one, allowing the same pretrained model to act as both drafter and verifier. S2D2 inserts a speculative verification step into standard block-diffusion decoding and uses lightweight routing policies to decide when verification is worth its cost. This yields a hybrid decoding trajectory in which diffusion proposes tokens in parallel, while the autoregressive mode acts as a local sequence-level critic. Across three mainstream block-diffusion families, S2D2 consistently improves the accuracy-speed tradeoff over strong confidence-thresholding baselines. On SDAR, we observe up to $4.7\times$ speedup over autoregressive decoding, and up to $1.57\times$ over a tuned dynamic decoding baseline while improving accuracy by up to $4.5$ points. On LLaDA2.1-Mini, S2D2 remains complementary to built-in self-correction, including a conservative setting where it is $4.4\times$ faster than the static baseline with slightly higher accuracy.

Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens from a fully masked sequence. Their standard confidence-based unmasking strategy systematically defers high-entropy logical connective tokens, degrading reasoning performance. We introduce LogicDiff, an inference-time method that replaces confidence-based unmasking with logic-role-guided unmasking. A lightweight classification head (4.2M parameters, 0.05% of the base model) predicts the logical role of each masked position (premise, connective, derived step, conclusion, or filler) from the base model's hidden states with 98.4% accuracy, and a dependency-ordered scheduler unmasks tokens in logical order. In zero-shot settings, LogicDiff improves LLaDA-8B-Instruct accuracy from 22.0% to 60.7% on GSM8K (+38.7 percentage points) and from 23.6% to 29.2% on MATH-500 (+5.6 pp), with less than 6% speed overhead. However, with 8-shot chain-of-thought prompting, the baseline reaches approximately 70% and LogicDiff provides no additional improvement. Analysis reveals that few-shot prompting implicitly resolves the same ordering problem that LogicDiff explicitly addresses, and that fixed role-based ordering can cause premature commitment to numerical values before sufficient context is available. Our results characterize the Flexibility Trap as primarily a zero-shot phenomenon and identify context-adaptive ordering as a key direction for future work.

Optical character recognition (OCR) has evolved from line-level transcription to structured document parsing, requiring models to recover long-form sequences containing layout, tables, and formulas. Despite recent advances in vision-language models, most existing systems rely on autoregressive decoding, which introduces sequential latency and amplifies error propagation in long documents. In this work, we revisit document OCR from an inverse rendering perspective, arguing that left-to-right causal generation is an artifact of serialization rather than an intrinsic property of the task. Motivated by this insight, we propose MinerU-Diffusion, a unified diffusion-based framework that replaces autoregressive sequential decoding with parallel diffusion denoising under visual conditioning. MinerU-Diffusion employs a block-wise diffusion decoder and an uncertainty-driven curriculum learning strategy to enable stable training and efficient long-sequence inference. Extensive experiments demonstrate that MinerU-Diffusion consistently improves robustness while achieving up to 3.2x faster decoding compared to autoregressive baselines. Evaluations on the proposed Semantic Shuffle benchmark further confirm its reduced dependence on linguistic priors and stronger visual OCR capability.

Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive (AR) models for language modeling, allowing flexible generation order and parallel generation of multiple tokens. However, this flexibility introduces a challenge absent in AR models: the decoding strategy -- which determines the order and number of tokens generated at each iteration -- critically affects sampling efficiency. Among decoding strategies explored in practice, confidence-based methods, which adaptively select which and how many tokens to unmask based on prediction confidence, have shown strong empirical performance. Despite this success, our theoretical understanding of confidence-based decoding remains limited. In this work, we develop the first theoretical analysis framework for confidence-based decoding in DLMs. We focus on an entropy sum-based strategy that continues unmasking tokens within each iteration until the cumulative entropy exceeds a threshold, and show that it achieves $\varepsilon$-accurate sampling in KL divergence with an expected number of iterations $\widetilde O(H(X_0)/\varepsilon)$, where $H(X_0)$ denotes the entropy of the target data distribution. Notably, this strategy yields substantial sampling acceleration when the data distribution has low entropy relative to the sequence length, while automatically adapting to the intrinsic complexity of data without requiring prior knowledge or hyperparameter tuning. Overall, our results provide a theoretical foundation for confidence-based decoding and may inform the design of more efficient decoding strategies for DLMs.

Discrete diffusion models are promising alternatives to autoregressive approaches for text generation, yet their decoding methods remain under-studied. Standard autoregressive search procedures, such as beam search, do not directly apply to iterative denoising, where hypotheses are complete intermediate sequences rather than left-to-right prefixes. Furthermore, existing diffusion decoding procedures only provide limited control over the diversity and coverage of retained hypotheses. In this work, we introduce D5P4, a beam-style decoding method tailored to discrete diffusion models, which casts intermediate beam selection as MAP inference under a partitioned Determinantal Point Process. This yields a model-internal batch objective that balances quality and diversity without external verifiers. Experiments on open-ended generation, question answering, and mathematical reasoning show that D5P4 improves diversity and pass@$k$ coverage while matching or surpassing baseline quality and fidelity

Auto-regressive models (ARMs) have established a dominant paradigm in language modeling. However, their strictly sequential sampling paradigm imposes fundamental constraints on both inference efficiency and modeling flexibility. To address these limitations, diffusion-based large language models (dLLMs) have been proposed, offering the potential for parallel sampling and flexible language modeling. Despite these advantages, current dLLMs sampling strategies rely primarily on token level information, which fails to account for global sequence structure and often yields suboptimal results. In this paper, we study the sampling order selection problem from the perspective of log-likelihood maximization. We show that this problem is NP-hard and propose an optimal sampling-rank-based approximation that makes the objective computationally tractable. We further prove that the tractable objective is optimized by sampling tokens in descending order of their attention-matrix column sums. This finding provides a principled justification for attention-guided sampling and offers a theoretically grounded alternative to greedy search. We instantiate this theoretical insight in a new training-free sampling algorithm, termed Attn-Sampler, and further propose dynamic attention thresholding for practical acceleration. Extensive experiments across multiple benchmarks validate the effectiveness of our proposed method, demonstrating that it achieves superior generation quality while enhancing the sampling parallelism.

Masked diffusion language models (MDLMs) have recently emerged as a new paradigm in language modeling, offering flexible generation dynamics and enabling efficient parallel decoding. However, existing decoding strategies for pre-trained MDLMs predominantly rely on token-level uncertainty criteria, while largely overlooking sequence-level information and inter-token dependencies. To address this limitation, we propose Dependency-Oriented Sampler (DOS), a training-free decoding strategy that leverages inter-token dependencies to inform token updates during generation. Specifically, DOS exploits attention matrices from transformer blocks to approximate inter-token dependencies, emphasizing information from unmasked tokens when updating masked positions. Empirical results demonstrate that DOS consistently achieves superior performance on both code generation and mathematical reasoning tasks. Moreover, DOS can be seamlessly integrated with existing parallel sampling methods, leading to improved generation efficiency without sacrificing generation quality.

Parallel decoding for Diffusion LLMs (dLLMs) is difficult because each denoising step provides only token-wise marginal distributions, while unmasking multiple tokens simultaneously requires accounting for inter-token dependencies. We propose Dependency-Aware Parallel Decoding (DAPD), a simple, training-free decoding method that uses self-attention to induce a conditional dependency graph over masked tokens. At each iteration, edges in this graph capture strong token interactions, while non-edges indicate weak dependence. Parallel decoding is then reduced to selecting an independent set on the graph and unmasking the selected tokens in parallel. This avoids co-updating strongly coupled tokens without auxiliary models or retraining. Experiments on LLaDA and Dream show that DAPD improves the accuracy-steps trade-off over existing methods and enables more globally distributed parallel updates that better exploit the any-order generation capability of dLLMs. The project is available at https://ai-isl.github.io/dapd

Masked diffusion language models enable parallel token decoding, providing a promising alternative to the sequential nature of autoregressive generation. However, their iterative denoising process remains computationally expensive because it repeatedly processes the entire sequence at every step. We observe that across these diffusion steps, most token representations remain stable; only a small subset, which we term salient tokens, contributes meaningfully to the next update. Leveraging this temporal sparsity, we present DyLLM, a training-free inference framework that accelerates decoding by selectively computing only these salient tokens. DyLLM identifies saliency by measuring the cosine similarity of attention contexts between adjacent denoising steps. It recomputes feed-forward and attention operations only for salient tokens while reusing cached activations for the remainder. Across diverse reasoning and code-generation benchmarks, DyLLM achieves up to 9.6x higher throughput while largely preserving the baseline accuracy of representative open-source diffusion LLMs, LLaDA, and Dream.

Unlike autoregressive language models, which terminate variable-length generation upon predicting an End-of-Sequence (EoS) token, Diffusion Language Models (DLMs) operate over a fixed maximum-length context window for a predetermined number of denoising steps. However, this process is independent of the required response length, resulting in computational waste for the majority of short responses common in reasoning and chat tasks. To address this problem, we conjecture that the latent prompt representation contains sufficient information to estimate the required output length. We provide empirical evidence for this phenomenon and propose a zero-shot mechanism to dynamically crop the context window before generation begins, leading to fewer diffusion steps and substantial computational savings. We evaluate our approach on four benchmarks with diverse tasks -- GSM8K (reasoning), HumanEval (code generation), IfEval (instruction following), and LongFormQA (question answering) -- revealing massive efficiency gains at minimal performance impact. We report significant reductions in FLOPs across all tasks, with no statistically significant performance degradation, and significant performance improvements in 2 out of 4 tasks.

Diffusion Language Models (DLMs) promise highly parallel text generation, yet their practical inference speed is often bottlenecked by suboptimal decoding schedulers. Standard approaches rely on 'scattered acceptance'-committing high confidence tokens at disjoint positions throughout the sequence. This approach inadvertently fractures the Key-Value (KV) cache, destroys memory locality, and forces the model into costly, repeated repairs across unstable token boundaries. To resolve this, we present the Longest Stable Prefix (LSP) scheduler, a training-free and model-agnostic inference paradigm based on monolithic prefix absorption. In each denoising step, LSP evaluates token stability via a single forward pass, dynamically identifies a contiguous left-aligned block of stable predictions, and snaps its boundary to natural linguistic or structural delimiters before an atomic commitment. This prefix-first topology yields dual benefits: systemically, it converts fragmented KV cache updates into efficient, contiguous appends; algorithmically, it preserves bidirectional lookahead over a geometrically shrinking active suffix, drastically reducing token flip rates and denoiser calls. Extensive evaluations on LLaDA-8B and Dream-7B demonstrate that LSP accelerates inference by up to 3.4x across rigorous benchmarks including mathematical reasoning, code generation, multilingual (CJK) tasks, and creative writing while matching or slightly improving output quality. By fundamentally restructuring the commitment topology, LSP bridges the gap between the theoretical parallelism of DLMs and practical hardware efficiency.

Diffusion language models generate text through iterative denoising under a uniform refinement rule applied to all tokens. However, tokens stabilize at different rates in practice, leading to substantial redundant refinement and motivating refinement control over the denoising process. Existing approaches typically assess refinement necessity from instantaneous, step-level signals under a fixed decoding process. In contrast, whether a token has converged is defined by how its prediction changes along its future refinement trajectory. Moreover, changing the refinement rule reshapes future refinement trajectories, which in turn determine how refinement rules should be formulated, making refinement control inherently dynamic. We propose Progressive Refinement Regulation (PRR), a progressive, trajectory-grounded refinement control framework that derives a token-level notion of empirical convergence progress from full decoding rollouts. Based on this signal, PRR learns a lightweight token-wise controller to regulate refinement via temperature-based distribution shaping under a progressive self-evolving training scheme. Experiments show that PRR substantially accelerates diffusion language model decoding while preserving generation quality.

Block diffusion enables efficient parallel refinement in diffusion language models, but its decoding behavior depends critically on block size. Existing block-sizing strategies rely on fixed rules or heuristic signals and do not account for the dependency geometry that determines which tokens can be safely refined together. This motivates a geometry view of diffusion decoding: regions with strong causal ordering require sequential updates, whereas semantically cohesive regions admit parallel refinement. We introduce GeoBlock, a geometry-aware block inference framework that determines block granularity directly from attention-derived dependency geometry. Instead of relying on predefined schedules or local confidence heuristics, GeoBlock analyzes cross-token dependency patterns to identify geometrically stable refinement regions and dynamically determines appropriate block boundaries during decoding. By adapting block granularity to the dependency geometry, GeoBlock preserves the parallel efficiency of block diffusion while enforcing dependency-consistent refinement that exhibits autoregressive reliability. GeoBlock requires no additional training and integrates seamlessly into existing block diffusion architectures. Extensive experiments across multiple benchmarks show that GeoBlock reliably identifies geometry-consistent block boundaries and improves the accuracy of block diffusion with only a small additional computational budget.

Diffusion Large Language Models (dLLMs) have emerged as a promising paradigm for parallel token generation, with block-wise variants garnering significant research interest. Despite their potential, existing dLLMs typically suffer from a rigid accuracy-parallelism trade-off: increasing the number of tokens per forward (TPF) via aggressive parallel decoding often leads to performance degradation and increased generation instability. We identify that this limitation stems from the model's inability to navigate high-parallelism regimes where approximation errors and local corruptions accumulate, ultimately undermining the reliability of parallel generation. To address this, we propose LightningRL, a post-training framework designed to directly optimize the speed-quality Pareto frontier of pre-trained dLLMs. Instead of forcing uniform parallelization, our approach leverages reinforcement learning to identify and reinforce high-parallelism trajectories that maintain generation accuracy. Built upon the Group Relative Policy Optimization (GRPO) framework, LightningRL introduces several enhancements tailored for dLLMs: (1) stabilized training via per-reward decoupled normalization; (2) token-level negative log-likelihood (NLL) regularization on correct trajectories to anchor model performance; and (3) a dynamic sampling strategy with TPF-aware filtering to enhance training efficiency. Experimental results across mathematical and coding benchmarks demonstrate that LightningRL consistently advances the Pareto frontier, achieving competitive task accuracy while significantly increasing parallelism, reaching an average TPF of 7.32 (with a peak of 11.10 on the MBPP dataset). Our code is available at https://github.com/SJTU-DENG-Lab/LightningRL.

Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive (AR) models, offering sub-linear generation latency and bidirectional capabilities that are particularly appealing for code generation and editing. Achieving sub-linear latency in discrete DLMs requires predicting multiple tokens in parallel. However, standard DLMs sample tokens independently from conditional marginal distributions, failing to capture the joint dependencies among concurrently generated tokens. As a result, they often lead to syntactic inconsistencies and break multi-token structures. In this work, we introduce CoDiLA (Coherent Diffusion with Local Autoregression), a method that reconciles parallel sampling with local dependency modeling. Rather than forcing the DLM to resolve fine-grained syntax, CoDiLA delegates local decoding to a small, auxiliary AR model operating on the diffusion latents. This design allows for parallel generation while ensuring sequential validity within a block and maintaining core DLM capabilities, including bidirectional modeling across blocks. We demonstrate that using a highly compact auxiliary AR model (e.g., 0.6B parameters) effectively eliminates coherence artifacts, establishing a new Pareto frontier for accuracy and speed in code generation benchmarks.

We study why continuous diffusion language models (DLMs) have lagged behind discrete diffusion approaches despite their appealing continuous generative dynamics. Under a controlled token--recovery study, we identify token rounding, the final projection from denoised embeddings to tokens, as a primary bottleneck. Building on these insights, we propose CoDAR (Continuous Diffusion with Contextual AutoRegressive Decoder), a two--stage framework that keeps diffusion entirely continuous in an embedding space while learning a strong, context--conditional discretizer: an autoregressive Transformer decoder that cross--attends to the denoised embedding sequence and performs contextualized rounding to tokens. Experiments on LM1B and OpenWebText demonstrate that CoDAR substantially improves generation quality over latent diffusion and becomes competitive with strong discrete DLMs, while exposing a simple decoder--temperature knob to navigate the fluency--diversity trade off.

Masked diffusion models (MDMs) generate text by iteratively selecting positions to unmask and then predicting tokens at those positions. Yet MDMs lack proper likelihood evaluation: the evidence lower bound (ELBO) is not only a loose bound on log-likelihood, but, as we show, is also computed under the training distribution rather than the test-time distribution. We resolve this within our DUEL framework, which unifies leading MDM sampling strategies that employ $\textit{deterministic}$ position selection. We prove that DUEL samplers admit $\textbf{exact likelihood computation under the test-time distribution}$ -- giving MDMs $\textit{proper}$ likelihood, and hence proper perplexity, for the first time. This proper perplexity is the natural analogue of autoregressive perplexity and lets us revisit key questions about MDMs. $\textbf{MDMs are substantially better than previously thought}$: the MDM-autoregressive perplexity gap shrinks by up to $32\%$ on in-domain data and $82\%$ on zero-shot benchmarks. DUEL enables the first principled comparison of fast,parallel samplers across compute budgets -- an analysis impossible with the ELBO and unreliable with generative perplexity -- identifying a strong default method. Finally, oracle search over position orderings reveals MDMs can far surpass autoregressive models -- achieving $36.47$ vs. $52.11$ perplexity on AG News -- demonstrating the ceiling of MDM performance has not yet been reached.

We present LLaDA-o, an effective and length-adaptive omni diffusion model for multimodal understanding and generation. LLaDA-o is built on a Mixture of Diffusion (MoD) framework that decouples discrete masked diffusion for text understanding and continuous diffusion for visual generation, while coupling them through a shared, simple, and efficient attention backbone that reduces redundant computation for fixed conditions. Building on MoD, we further introduce a data-centric length adaptation strategy that enables flexible-length decoding in multimodal settings without architectural changes. Extensive experiments show that LLaDA-o achieves state-of-the-art performance among omni-diffusion models on multimodal understanding and generation benchmarks, and reaches 87.04 on DPG-Bench for text-to-image generation, supporting the effectiveness of unified omni diffusion modeling. Code is available at https://github.com/ML-GSAI/LLaDA-o.

Masked discrete diffusion models (MDMs) are a promising new approach to generative modelling, offering the ability for parallel token generation and therefore greater efficiency than autoregressive counterparts. However, achieving an optimal balance between parallel generation and sample quality remains an open problem. Current approaches primarily address this issue through fixed, heuristic parallel sampling methods. There exist some recent learning based approaches to this problem, but its formulation from the perspective of variational inference remains underexplored. In this work, we propose a variational inference framework for learning parallel generation orders for MDMs. As part of our method, we propose a parameterisation for the approximate posterior of generation orders which facilitates parallelism and efficient sampling during training. Using this method, we conduct preliminary experiments on the GSM8K dataset, where our method performs competitively against heuristic sampling strategies in the regime of highly parallel generation. For example, our method achieves 33.1% accuracy with an average of only only 4 generation steps, compared to 23.7-29.0% accuracy achieved by standard competitor methods in the same number of steps. We believe further experiments and analysis of the method will yield valuable insights into the problem of parallel generation with MDMs.

Diffusion-based large language models (dLLMs) have shown promising performance across various reasoning tasks, establishing themselves as an alternative to autoregressive large language models (LLMs). Unlike autoregressive LLMs that generate one token per step based on all previous tokens, dLLMs theoretically enable parallel generation of multiple tokens at each decoding step. However, recent dLLMs still favor one-token-per-step generation in practice, as directly decoding multiple masked tokens often leads to degraded generation quality and stability. This reveals a substantial gap between the theoretical parallelism and practical performance of dLLMs. To bridge this gap, we introduce an adaptive parallel decoding approach, namely DiCo, which features a three-phase divide-and-conquer paradigm to unleash the inherent parallelism of dLLMs. During the Divide phase, DiCo first explores the input masked sequence and identifies masked tokens as seed tokens, which are then expanded to construct a set of local clusters. During the Conquer phase, DiCo performs parallel decoding across different local clusters constructed in the Divide phase. The divide-and-conquer process repeatedly alternates between the Divide and Conquer phases until convergence. During the Finalize phase, DiCo decodes the remaining few masked tokens using an effective fine-grained compound decoding scheme to finalize the generation. Extensive experiments demonstrate that DiCo can achieve significant inference speedups while maintaining competitive generation quality.

Diffusion Language Models (DLMs) are often advertised as enabling parallel token generation, yet practical fast DLMs frequently converge to left-to-right, autoregressive (AR)-like decoding dynamics. In contrast, genuinely non-AR generation is promising because it removes AR's sequential bottleneck, better exploiting parallel hardware to reduce synchronization/communication overhead and improve latency scaling with output length. We argue that a primary driver of AR-like decoding is a mismatch between DLM objectives and the highly sequential structure of widely used training data, including standard pretraining corpora and long chain-of-thought (CoT) supervision. Motivated by this diagnosis, we propose NAP (Non-Autoregressive Parallel DLMs), a proof-of-concept, data-centric approach that better aligns supervision with non-AR parallel decoding. NAP curates examples as multiple independent reasoning trajectories and couples them with a parallel-forced decoding strategy that encourages multi-token parallel updates. Across math reasoning benchmarks, NAP yields stronger performance under parallel decoding than DLMs trained on standard long CoT data, with gains growing as parallelism increases. Our results suggest that revisiting data and supervision is a principled direction for mitigating AR-like behavior and moving toward genuinely non-autoregressive parallel generation in DLMs. Our code is available at https://github.com/pixeli99/NAP.