Diffusion Language Model Inference

Diffusion language models (DLMs) expose their denoising trajectories, offering a natural handle for inference-time control; accordingly, an ideal hallucination mitigation framework should intervene during generation using this model-native signal rather than relying on an externally trained hallucination classifier. Toward this, we formulate commitment uncertainty localization: given a denoising trajectory, identify token positions whose cross-chain entropy exceeds an unsupervised threshold before factually unreliable commitments propagate into self-consistent but incorrect outputs. We introduce a suite of trajectory-level assessments, including a cross-chain divergence-at-hallucination (CDH) metric, for principled comparison of localization methods. We also introduce OSCAR, a training-free inference-time framework operationalizing this formulation. OSCAR runs N parallel denoising chains with randomized reveal orders, computes cross-chain Shannon entropy to detect high-uncertainty positions, and then performs targeted remasking conditioned on retrieved evidence. Ablations confirm that localization and correction contribute complementary gains, robust across N in {4, 8, 16}. On TriviaQA, HotpotQA, RAGTruth, and CommonsenseQA using LLaDA-8B and Dream-7B, OSCAR enhances generation quality by significantly reducing hallucinated content and improving factual accuracy through uncertainty-guided remasking, which also facilitates more effective integration of retrieved evidence. Its native entropy-based uncertainty signal surpasses that of specialized trained detectors, highlighting an inherent capacity of diffusion language models to identify factual uncertainty that is not present in the sequential token commitment structure of autoregressive models.

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-based large language models (dLLMs) rely on bidirectional attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum entropy of newly decoded token distributions as a constant-cost signal for deciding when to recompute. Our design is grounded in two empirical observations: (1) decoded token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature volatility of decoded tokens persists for multiple steps after unmasking, motivating recomputation of the $k$ most recently decoded tokens. The skip-or-recompute decision requires only $O(V)$ computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves $15.2\times$-$26.4\times$ speedup on standard benchmarks and $22.4\times$-$24.1\times$ on chain-of-thought benchmarks, with competitive accuracy and decision overhead accounting for only $0.5\%$ of inference time. Code is available at https://github.com/mscheong01/EntropyCache.

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.

Discrete Diffusion Language Models have emerged as a compelling paradigm for unified multimodal generation, yet their deployment is hindered by high inference latency arising from iterative decoding. Existing acceleration strategies often require expensive re-training or fail to leverage the 2D spatial redundancy inherent in visual data. To address this, we propose Locality-Aware Dynamic Rescue (LADR), a training-free method that expedites inference by exploiting the spatial Markov property of images. LADR prioritizes the recovery of tokens at the ''generation frontier'', regions spatially adjacent to observed pixels, thereby maximizing information gain. Specifically, our method integrates morphological neighbor identification to locate candidate tokens, employs a risk-bounded filtering mechanism to prevent error propagation, and utilizes manifold-consistent inverse scheduling to align the diffusion trajectory with the accelerated mask density. Extensive experiments on four text-to-image generation benchmarks demonstrate that our LADR achieves an approximate 4 x speedup over standard baselines. Remarkably, it maintains or even enhances generative fidelity, particularly in spatial reasoning tasks, offering a state-of-the-art trade-off between efficiency and quality.

Diffusion large language models (dLLMs) are emerging as a promising alternative to autoregressive models (ARMs) due to their ability to capture bidirectional context and the potential for parallel generation. Despite the advantages, dLLM inference remains computationally expensive as the full input context is processed at every iteration. In this work, we analyze the generation dynamics of dLLMs and find that intermediate representations, including key, value, and hidden states, change only subtly across successive iterations. Leveraging this insight, we propose ES-dLLM, a training-free inference acceleration framework for dLLM that reduces computation by skipping tokens in early layers based on the estimated importance. Token importance is computed with intermediate tensor variation and confidence scores of previous iterations. Experiments on LLaDA-8B and Dream-7B demonstrate that ES-dLLM achieves throughput of up to 226.57 and 308.51 tokens per second (TPS), respectively, on an NVIDIA H200 GPU, delivering 5.6$\times$ to 16.8$\times$ speedup over the vanilla implementation and up to 1.85$\times$ over the state-of-the-art caching method, while preserving generation quality.

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.

Reinforcement Learning with Verifiable Rewards (RLVR) successfully aligns autoregressive models in rigorous domains like mathematical reasoning and code generation. However, applying RLVR to Diffusion Large Language Models (dLLMs) is hindered by intractable sequence-level likelihoods, forcing existing methods to rely on unstable, high-variance approximations. We propose Sequence Likelihood-Free Policy Optimization (LFPO), an efficient framework that circumvents this bottleneck by formulating policy alignment as discrete velocity alignment. Inspired by continuous flow matching, LFPO eschews trajectory-level estimation and directly optimizes denoising logits via contrastive token-level updates, providing a stable, low-variance surrogate. Furthermore, predicting terminal solutions from intermediate states implicitly straightens the generation trajectory, enabling high-quality sampling with significantly fewer denoising iterations. Extensive experiments on reasoning and coding benchmarks demonstrate that LFPO matches or exceeds the accuracy of likelihood-based baselines while substantially accelerating inference.

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 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.

Although diffusion language models (DLMs) are evolving quickly, many recent models converge on a set of shared components. These components, however, are distributed across ad-hoc research codebases or lack transparent implementations, making them difficult to reproduce or extend. As the field accelerates, there is a clear need for a unified framework that standardizes these common components while remaining flexible enough to support new methods and architectures. To address this gap, we introduce dLLM, an open-source framework that unifies the core components of diffusion language modeling -- training, inference, and evaluation -- and makes them easy to customize for new designs. With dLLM, users can reproduce, finetune, deploy, and evaluate open-source large DLMs such as LLaDA and Dream through a standardized pipeline. The framework also provides minimal, reproducible recipes for building small DLMs from scratch with accessible compute, including converting any BERT-style encoder or autoregressive LM into a DLM. We also release the checkpoints of these small DLMs to make DLMs more accessible and accelerate future research.

Diffusion language models (DLMs) have recently emerged as a promising alternative to autoregressive (AR) approaches, enabling parallel token generation beyond a rigid left-to-right order. Despite growing empirical success, the theoretical understanding of how unmasking schedules -- which specify the order and size of unmasked tokens during sampling -- affect generation quality remains limited. In this work, we introduce a distribution-agnostic unmasking schedule for DLMs that adapts to the (unknown) dependence structure of the target data distribution, without requiring any prior knowledge or hyperparameter tuning. In contrast to prior deterministic procedures that fix unmasking sizes, our method randomizes the number of tokens revealed at each iteration. We show that, for two specific parameter choices, the sampling convergence guarantees -- measured by Kullback-Leibler (KL) divergence -- scale as $\widetilde O(\mathsf{TC}/K)$ and $\widetilde O(\mathsf{DTC}/K)$ respectively. Here, $K$ is the number of iterations, and $\mathsf{TC}$ and $\mathsf{DTC}$ are the total correlation and dual total correlation of the target distribution, capturing the intrinsic dependence structure underlying the data. Importantly, our guarantees hold in the practically relevant parallel-sampling regime $K<L$ where $L$ is the token sequence length. These results significantly improve upon prior convergence theories and yield substantial sampling acceleration for low-complexity distributions. Overall, our findings unveil the adaptivity of DLMs to intrinsic data structures and shed light on the benefit of randomized unmasking sizes in inference schedule design.

Diffusion Language Models (DLMs) have recently achieved strong results in text generation. However, their multi-step sampling leads to slow inference, limiting practical use. To address this, we extend Inverse Distillation, a technique originally developed to accelerate continuous diffusion models, to the discrete setting. Nonetheless, this extension introduces both theoretical and practical challenges. From a theoretical perspective, the inverse distillation objective lacks uniqueness guarantees, which may lead to suboptimal solutions. From a practical standpoint, backpropagation in the discrete space is non-trivial and often unstable. To overcome these challenges, we first provide a theoretical result demonstrating that our inverse formulation admits a unique solution, thereby ensuring valid optimization. We then introduce gradient-stable relaxations to support effective training. As a result, experiments on multiple DLMs show that our method, Inverse-distilled Diffusion Language Models (IDLM), reduces the number of inference steps by 4x-64x, while preserving the teacher model's generation quality. We provide the code, model checkpoints, and video tutorials on the project page: https://david-cripto.github.io/idlm-project-page

We introduce Evo, a duality latent trajectory model that bridges autoregressive (AR) and diffusion-based language generation within a continuous evolutionary generative framework. Rather than treating AR decoding and diffusion generation as separate paradigms, Evo reconceptualizes text generation as a latent flow: each token is associated with a vector-valued embedding that evolves over a progression variable $t_i \in [0, 1]$, indicating its semantic maturity. Low $t_i$ values correspond to confident AR-like refinement, while high values invoke diffusion-style planning, allowing the model to adaptively balance AR and diffusion based on uncertainty. Theoretically, we show that both AR and diffusion models emerge as discretizations of a shared probability flow, and we derive Evo's training objective from a unified variational ELBO. The model is implemented as a time-conditioned Transformer governed by a shared vector field, trained end-to-end to jointly infer latent codes and their progression times. During decoding, Evo performs efficient, semantics-aware refinement, achieving high-quality outputs without sacrificing speed. Empirically, Evo 8B achieves state-of-the-art or highly competitive results on 15 diverse benchmarks, including reasoning (GSM8K, ARC-C), code generation (HumanEval, MBPP), and general language understanding, while maintaining fast inference speed. Our results demonstrate that Evo delivers a new paradigm for LLM design with strong generation quality, robust symbolic reasoning, and decoding efficiency.

Diffusion Language Models (DLMs) incur high inference cost due to iterative denoising, motivating efficient pruning. Existing pruning heuristics largely inherited from autoregressive (AR) LLMs, typically preserve attention sink tokens because AR sinks serve as stable global anchors. We show that this assumption does not hold for DLMs: the attention-sink position exhibits substantially higher variance over the full generation trajectory (measured by how the dominant sink locations shift across timesteps), indicating that sinks are often transient and less structurally essential than in AR models. Based on this observation, we propose ${\bf \texttt{Sink-Aware Pruning}}$, which automatically identifies and prunes unstable sinks in DLMs (prior studies usually keep sinks for AR LLMs). Without retraining, our method achieves a better quality-efficiency trade-off and outperforms strong prior pruning baselines under matched compute. Our code is available at https://github.com/VILA-Lab/Sink-Aware-Pruning.

Diffusion language models are a promising alternative to autoregressive models due to their potential for faster generation. Among discrete diffusion approaches, Masked diffusion currently dominates, largely driven by strong perplexity on language modeling benchmarks. In this work, we present the first scaling law study of uniform-state and interpolating discrete diffusion methods. We also show that Masked diffusion models can be made approximately 12% more FLOPs-efficient when trained with a simple cross-entropy objective. We find that perplexity is informative within a diffusion family but can be misleading across families, where models with worse likelihood scaling may be preferable due to faster and more practical sampling, as reflected by the speed-quality Pareto frontier. These results challenge the view that Masked diffusion is categorically the future of diffusion language modeling and that perplexity alone suffices for cross-algorithm comparison. Scaling all methods to 1.7B parameters, we show that uniform-state diffusion remains competitive on likelihood-based benchmarks and outperforms autoregressive and Masked diffusion models on GSM8K, despite worse validation perplexity. We provide the code, model checkpoints, and video tutorials on the project page: http://s-sahoo.github.io/scaling-dllms

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.

Discrete Diffusion Language Models (DLMs) offer a promising non-autoregressive alternative for text generation, yet effective mechanisms for inference-time control remain relatively underexplored. Existing approaches include sampling-level guidance or trajectory optimization mechanisms. In this work, we study the paradigm of reference-based latent steering for DLMs. We introduce Iterative Latent Representation Refinement (ILRR), an efficient framework for steering DLMs using a reference text as a high-level semantic blueprint. ILRR extracts and injects reference-derived semantic signals into the evolving activations of the generated sequence, enabling tunable transfer of coarse properties such as sentiment. We further introduce Spatially Modulated Steering, an extension that enables long-form generation to be guided by shorter references by regulating intensity across the sequence. Empirically, we demonstrate that ILRR achieves effective control on LLaDA and MDLM architectures with low computational overhead, requiring only one additional parallel forward pass per denoising step. Under comparable compute budgets, ILRR improves attribute accuracy over baselines by 10% to 60% points. Our results suggest that the iterative, global denoising process makes DLMs a natural substrate for effective sequence-wide activation-level control.

Diffusion language models (DLMs) generate text through iterative denoising, but inference requires full-sequence attention at every iteration, resulting in substantial redundant computation on masked tokens. Block-wise diffusion can reduce this cost, yet it typically relies on retraining and constrained update orders, limiting its direct applicability to pretrained DLMs. Our token-level analysis reveals pronounced structural locality in DLM inference. Decoding is driven by a small set of prefix-localized active tokens; the influence of distant undecoded context diminishes rapidly, and decoded tokens exhibit stage-wise temporal stability, enabling reuse of intermediate representations except for a brief post-decode transient. Motivated by these observations, we propose \placeholder\footnote{The source code is available at https://github.com/vhicrgit/Window-Diffusion.}, a window-based token pruning and caching method for inference. We maintain a local computation window that slides rightward as denoising progresses, and partition undecoded tokens into: (i) active tokens that are computed online, (ii) buffer tokens whose KV states are cached and periodically refreshed, and (iii) far-field tokens that are pruned outside the window. Computation is restricted to active and buffer tokens within the window, while far-field tokens are omitted at each stage. Experiments on LLaDA and Dream show that, under matched compute budgets, our method achieves up to $99\times$ inference speedup while largely preserving generation performance.

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.

Uniform-noise discrete diffusion and flow models generate sequences non-autoregressively through iterative, context-dependent token replacements. However, these models are typically formulated as time-inhomogeneous continuous- or discrete-time Markov chains (CTMC/DTMC), sampled using independent Bernoulli change decisions per discretization step. This induces Poisson-binomial variance in per-position jump counts that grows with the number of required edits, leading to the common under-editing (residual noise) and over-editing (cascading substitutions) failure modes that degrade sample quality. We identify this sampler-induced variance as an orthogonal source of degradation, distinct from model-side errors and addressable purely at inference time. We propose Systematic Hazard Sampling (SHS), a training-free, drop-in, and hyperparameter-free inference principle for any sampler that admits a stay-vs.-replace decomposition. SHS models per-token edits as events driven by cumulative hazard (CTMC) or jump mass (DTMC) and triggers an edit whenever this quantity exceeds unit-spaced thresholds with a single random phase per position. For any fixed cumulative mass, this preserves the expected jump count while achieving the minimum conditional variance possible among unbiased integer estimators (at most 1/4), without altering per-jump destination sampling. Experiments on four uniform-noise discrete diffusion and flow language models spanning ~110M to ~3B parameters show that SHS consistently improves sample quality across numbers of function evaluations (NFE). Code is available at https://github.com/Jang-seunghwan/Systematic-Hazard-Sampling.

Masked Discrete Diffusion Models (MDMs) have achieved strong performance across a wide range of multimodal tasks, including image understanding, generation, and editing. However, their inference speed remains suboptimal due to the need to repeatedly process redundant masked tokens at every sampling step. In this work, we propose Sparse-LaViDa, a novel modeling framework that dynamically truncates unnecessary masked tokens at each inference step to accelerate MDM sampling. To preserve generation quality, we introduce specialized register tokens that serve as compact representations for the truncated tokens. Furthermore, to ensure consistency between training and inference, we design a specialized attention mask that faithfully matches the truncated sampling procedure during training. Built upon the state-of-the-art unified MDM LaViDa-O, Sparse-LaViDa achieves up to a 2x speedup across diverse tasks including text-to-image generation, image editing, and mathematical reasoning, while maintaining generation quality.

Discrete diffusion-based multimodal large language models (dMLLMs) have emerged as a promising alternative to autoregressive MLLMs thanks to their advantages in parallel decoding and bidirectional context modeling, but most existing dMLLMs incur significant computational overhead during inference due to the full-sequence attention computation in each denoising step. Pioneering studies attempt to resolve this issue from a modality-agnostic perspective via key-value cache optimization or efficient sampling but most of them overlook modality-specific visual token redundancy. In this work, we conduct a comprehensive study on how visual token redundancy evolves with different dMLLM architectures and tasks and how visual token pruning affects dMLLM responses and efficiency. Specifically, our study reveals that visual redundancy emerges only in from-scratch dMLLMs while handling long-answer tasks. In addition, we validate that visual token pruning introduces non-negligible information loss in dMLLMs and only from-scratch dMLLMs can recover the lost information progressively during late denoising steps. Furthermore, our study shows that layer-skipping is promising for accelerating AR-to-diffusion dMLLMs, whereas progressive or late-step pruning is more effective for from-scratch dMLLMs. Overall, this work offers a new perspective on efficiency optimization for dMLLMs, greatly advancing their applicability across various multimodal understanding tasks. The code is available at: https://github.com/Yrdal3910/dMLLM-Visual-Token-Redundancy-Analysis.

While most autoregressive LLMs are constrained to one-by-one decoding, diffusion LLMs (dLLMs) have attracted growing interest for their potential to dramatically accelerate inference through parallel decoding. Despite this promise, the conditional independence assumption in dLLMs causes parallel decoding to ignore token dependencies, inevitably degrading generation quality when these dependencies are strong. However, existing works largely overlook these inherent challenges, and evaluations on standard benchmarks (e.g., math and coding) are not sufficient to capture the quality degradation caused by parallel decoding. To address this gap, we first provide an information-theoretic analysis of parallel decoding. We then conduct case studies on analytically tractable synthetic list operations from both data distribution and decoding strategy perspectives, offering quantitative insights that highlight the fundamental limitations of parallel decoding. Building on these insights, we propose ParallelBench, the first benchmark specifically designed for dLLMs, featuring realistic tasks that are trivial for humans and autoregressive LLMs yet exceptionally challenging for dLLMs under parallel decoding. Using ParallelBench, we systematically analyze both dLLMs and autoregressive LLMs, revealing that: (i) dLLMs under parallel decoding can suffer dramatic quality degradation in real-world scenarios, and (ii) current parallel decoding strategies struggle to adapt their degree of parallelism based on task difficulty, thus failing to achieve meaningful speedup without compromising quality. Our findings underscore the pressing need for innovative decoding methods that can overcome the current speed-quality trade-off. We release our benchmark to help accelerate the development of truly efficient dLLMs.

Diffusion Large Language Models (dLLMs) have shown strong reasoning capabilities, yet further improving them typically requires costly post-training with additional data and supervision. We ask whether a post-trained dLLM can improve itself at inference time without additional training, data, or reward models. This requires a guidance signal from the checkpoints alone that is well-defined on the partially masked intermediate states of dLLMs, which existing methods fail to provide. Here we propose Reward-Free Guidance (RFG), a training-free framework for inference-time self-improvement of dLLMs that derives such a signal from the checkpoints themselves. We theoretically demonstrate that a reward signal for a partially masked state can be parameterized by the log-likelihood ratio between a post-trained policy dLLM and its reference model. In practice, this signal can be obtained using off-the-shelf checkpoints alone. Extensive experiments show that RFG consistently improves state-of-the-art post-trained dLLMs by up to 16.1%. Furthermore, despite being entirely training-free, RFG delivers gains that rival or even surpass resource-intensive reinforcement learning techniques.

Efficiently scaling Large Language Models (LLMs) necessitates exploring alternatives to dominant autoregressive (AR) methods, with Masked Diffusion Models (MDMs) emerging as candidates. However, comparing AR (typically decoder-only) and MDM (often encoder-only) paradigms is confounded by differing architectures, obscuring true algorithmic and efficiency trade-offs. This research decouples these factors by evaluating MDMs within a decoder-only framework to: (1) Equitably compare MDM (as Any-Order AR) and standard AR paradigms through discrepancies on orders. (2) Investigate MDM architectural impacts on computational efficiency. We show decoder-only MDMs, despite a larger modeling space, can achieve significant inference speedups ($\sim25\times$) and comparable perplexity with techniques like temperature annealing, offering a path to reduced inference compute. This work provides insights for developing more computationally efficient foundation models by disentangling core modeling choices from architectural influences. Code is available at https://github.com/scxue/AO-GPT-MDM.