Overview

Diffusion language models generate text by starting from corrupted or masked sequences and repeatedly refining them, using context from both earlier and later positions. Main approaches include discrete masked diffusion, continuous latent or flow-based models, and block or hybrid designs that combine parallel refinement with autoregressive structure. Work explores adaptive token ordering, remasking, distillation, and caching to improve reasoning, code generation, multimodal tasks, and inference speed. A central tension is balancing faster, more parallel decoding against token dependencies and output quality, while reliable evaluation and the cost of training from scratch remain open challenges.

Latest papers

494 papers69 papers in the last four weeks, down 10% on the four weeks before. 0.5% of all new papers.

Oct 1, 2026

Hierarchical Continuous Diffusion Language Models

Discrete diffusion language models offer a compelling alternative to autoregressive generation for tasks demanding bidirectional reasoning and global constraint satisfaction. Yet they share a structural bottleneck: when decoding in parallel, each token is sampled independently from its marginal, severing the statistical dependencies among the tokens decoded together. Continuous diffusion language models avoid this by denoising a shared continuous state, but their denoiser sees only that state, so nothing ties it to a valid token configuration until it is finally decoded. To address this, we propose Hierarchical Continuous Diffusion Language Models (HC-DLM), which couple discrete token generation with a continuous latent trajectory in a single, principled denoising process, whose training objective is derived from a variational bound on the token likelihood. In contrast to recent methods that attach continuous context to a self-contained discrete chain, HC-DLM makes the latent the only persistent generative state: tokens are read out from it at every step and feed back as a scaffold for the next latent update. On structured reasoning (Sudoku), mathematical planning (Countdown) and language modeling (LM1B), HC-DLM improves over discrete and continuous diffusion baselines at matched model size, in puzzle accuracy on Sudoku and Countdown and in generative perplexity on LM1B. Project page: https://hc-dlm.github.io/.
Oct 1, 2026

ITC-MoE: Importance-guided Token-aware Compression for MoE Diffusion Language Models

Mixture-of-Experts (MoE) Diffusion Language Models (DLMs) offer flexible parallel decoding and increased model capacity, but their large number of expert parameters incurs substantial computation and storage costs. Existing low-rank MoE compression methods largely rely on static factorization and fixed rank allocation, which overlook the distinctive properties of MoE DLMs. Specifically, we identify two properties: cross-mode non-uniform redundancy, where parameter redundancy and sensitivity to rank truncation vary across the input, output, and expert modes, and token-wise utilization variation, where hot and cold tokens exhibit distinct spectral characteristics and expert activation patterns. To address these challenges, we propose ITC-MoE, an Importance-guided Token-aware Compression framework for MoE DLMs. ITC-MoE consists of two complementary components. First, Importance-guided Adaptive Tucker Compression (IATC) incorporates activation and gradient importance into expert weight transformation, jointly factorizes expert weights across multiple modes, and adaptively allocates ranks under a fixed parameter budget. Second, Token-aware Compensation and Routing (TCR) applies lightweight low-rank compensation to compression-sensitive hot tokens and restricts the candidate expert set for cold tokens with concentrated routing patterns. By jointly adapting compression capacity and inference execution to both parameter redundancy and token-wise variation, ITC-MoE substantially reduces the computation and storage costs of MoE DLMs while preserving their generation quality. For example, on SDAR-30B-A3B-Chat-b32, ITC-MoE maintains an accuracy of 96.33% on MultiArith under a 30% compression budget, while achieving up to a 7.22x end-to-end speedup. The code is publicly available at https://github.com/lianjunl13-sudo/ITC-MoE.
Oct 1, 2026

Know When to Hold 'em: Correct-Token Retention in Uniform-State Diffusion Language Models

Uniform-state diffusion models (USDMs) can revise any token at any denoising step, which lets them correct their own mistakes, a key advantage over masked diffusion. Self-correction, however, requires both revising incorrect tokens and retaining correct ones, and we show that current USDMs lack the latter. Even under greedy-tail decoding, state-of-the-art USDMs (DUO, UDLM, and uniform-noise SEDD) keep revising 173--270 of 512 positions at every step, and these large, uncoordinated edits collapse sample diversity. A random-token corruption experiment traces this deficit to the models themselves: they reconstruct clean and corrupted tokens with nearly identical accuracy, even though clean tokens are easier targets. A decomposition of the validation NELBO shows that training barely rewards retention: incorrect predictions are heavily penalized at corrupted positions but almost free at clean ones. We propose Correct-Token Retention Regularization (CTR-Reg), a simple but effective auxiliary loss that trains the model to retain tokens left unperturbed by the forward process and requires no change to the sampler. CTR-Reg improves clean-token accuracy by 26.5 percentage points on average across six benchmarks, while leaving corrupted-token accuracy virtually unchanged, and its per-step revisions converge to only 3--11 positions. With just five greedy-tail steps, generative perplexity more than halves under CTR-Reg for all three models while diversity is preserved, and these gains hold across sampling budgets. Our results identify correct-token retention as a key missing ingredient for self-correcting diffusion language models, and demonstrate an effective fix.
Oct 1, 2026

Temporally-Resolved Token Attribution Reveals the Generation Dynamics of Diffusion Language Models

This work presents Diffusion Layer Integrated Gradients (DLIG), a token attribution method for diffusion language models (DLMs) that extends Integrated Gradients (IG~\cite{sundararajan2017axiomatic}) to arbitrary layers and denoising steps. DLIG attributes a DLM's progressive commitment to a self-generated or fixed completion for an input prompt. We establish direct correspondences between DLIG and the IG axioms of completeness, implementation invariance, linearity, and symmetry preservation. As a lightweight complement to interventional analysis, DLIG provides an inexpensive first check of mechanistic hypotheses across the denoising trajectory. We demonstrate this on word-sense disambiguation, multi-hop graph reasoning, and sentence infilling, revealing how DLMs draw on inputs across positions, layers, and denoising steps.
Oct 1, 2026

Scaling and Distilling Text Embeddings for Better Diffusibility

Diffusion language models (DLMs) offer a promising alternative to autoregressive (AR) language generation. Recent advances in continuous DLMs, which apply latent diffusion to continuous text embeddings, raise a practical question: which embedding makes the best latent space, i.e., the most diffusible? To answer this, we search through different embeddings and find that scaling the embedding model to stronger ones within the same family (T5 to T5Gemma-1 to T5Gemma-2) greatly improves generative performance. But the raw T5Gemma-2 embeddings are still not optimal. They are so discriminative that even the embeddings of plausible alternative words are separated, which makes the generation vulnerable to imperfect sampling. Consequently, continuous diffusion often fails to reach any of them and ends up at an invalid embedding instead. To address this, we distill T5Gemma-2 into a student encoder that learns the teacher's decoded probabilities as soft labels. Learning from such soft labels makes the student pull the alternative embeddings closer while maintaining the encoding-decoding mechanism. The distilled embeddings form a more connected and diffusible latent space, improving over the vanilla T5Gemma-2 embeddings. As a result, our medium-sized DLM achieves Gen. PPL 17.8 (against real-text PPL 15.4) at real-text entropy on OpenWebText, outperforming GPT-2-M on Gen. PPL.
Oct 1, 2026

Two Clocks in Diffusion MLLMs: When Answers Stabilize Before Rationales Unfold

An answer candidate in a masked diffusion MLLM can stabilize while its rationale is still unfolding. We distinguish retrospective stabilization of the logged candidate from token commitment, and examine these two clocks relative to rationale generation. Analyzing our results across three visual question-answering benchmarks, we find that 89.4-98.1% of the rationale-side canvas remains unwritten at stabilization in single-block, EOS-suppressed LaViDa runs. On VBench, reducing block length from 128 to 8 changes this fraction from 89.4% to 1.7%, together with answer coverage and the eligible observation window. Under EOS-enabled prompting, direct instructions improve Nemotron's overall accuracy by 15.0 and 19.5 percentage points on M3CoT and ScienceQA, but reduce LaViDa/VBench accuracy by 11.0 points. A symmetric decomposition associates the larger absolute component of each change with coverage rather than conditional accuracy. Matched-canvas image ablations measure visual sensitivity alongside answer stabilization, separating the two temporal readouts. Together, these measurements distinguish answer stabilization, rationale unfolding, and visual sensitivity, and identify coverage as the larger component of the prompting differences.
Oct 1, 2026

Clock Diffusion: Efficient Semi-Autoregressive Continuous Diffusion Language Models

Recent works on continuous diffusion for discrete data have demonstrated performance on par with comparable discrete diffusion models. However, these continuous counterparts lack key features that are essential to practical use as language models, namely variable-length generation and support for a key-value cache, and they still lag behind the frontier of autoregressive and discrete diffusion quality. In this work, we address these limitations. We do so by introducing a model parameterization that uses position-dependent noise schedules to define semi-autoregressive (SAR) continuous diffusion language models (DLMs). Together with efficient training and sampling algorithms, we call this framework Clock Diffusion, and we present two special cases of our method: block and sliding window generation. We then define ClockDLMs, a family of Gaussian DLMs based on sliding window Clock Diffusion that attain state-of-the-art diffusion likelihood bounds on OpenWebText, even beating the performant block SAR discrete diffusion models. ClockDLMs trained on TinyGSM also substantially outperform continuous baselines on the GSM8K benchmark and match and exceed comparable SAR discrete diffusion models. Finally, building on our parameterization, we propose more efficient samplers that we dub Cache Grab, which adapt techniques from accelerated inference in discrete diffusion, such as committing tokens whose probabilities exceed a confidence threshold and self-speculative decoding, further improving our models' quality and efficiency.
Sep 30, 2026

Exploring More, Reasoning Better: Stepwise Risk-Sensitive GRPO for Diffusion Language Models

Diffusion large language models (dLLMs) generate text by denoising a sequence or successive blocks, allowing several tokens to be revealed in parallel. Reinforcement learning with verifiable rewards (RLVR) reuses terminal feedback across these decisions, even as their conditioning context changes. We propose stepwise risk-sensitive GRPO (StepRS-GRPO), which varies the risk coefficient of the group-advantage transformation across denoising states while retaining the underlying trainer. For binary rewards, we show that this transformation is exactly a prompt- and state-dependent rescaling of centered outcome advantages. A capability-based calibration suggests a coefficient scale, while endpoint and interpolation ablations guide schedule selection. Across multiple dLLM backbones and mathematical reasoning benchmarks, StepRS-GRPO improves both pass@1 accuracy and pass@k coverage over centered GRPO, while increasing answer diversity. In our ablation studies, mass-matched controls support the contributions of state allocation and schedule direction, and the gains persist after matching the root mean square (RMS) of the advantages to that of centered GRPO. Reasoning-trace diagnostics further show that the diversity gains from StepRS-GRPO extend beyond final-answer strings.
Sep 30, 2026

Distribution Matching Distillation for Continuous Diffusion Language Models

Continuous diffusion language models generate all tokens in parallel, yet high-quality generation can still require hundreds of network evaluations (NFEs). We study how distributional distillation can reduce this cost by exploiting the student's probabilistic token outputs. Our unified formulation connects the student's output parameterization to the resulting gradient estimators and yields two methods with the same student architecture and reverse-KL matching objective: Simplex-DMD uses continuous token relaxations and pathwise gradients, while Reinforce-DMD uses categorical sampling and REINFORCE with a learned density ratio. We develop both methods for multi-step generation and investigate the training and sampling choices associated with each parameterization. On OpenWebText, for sequences of 1,024 tokens, Simplex-DMD achieves a generative perplexity of 45.6 at a unigram entropy of 5.44 nats in just 4 NFEs, a 49% reduction relative to the strongest evaluated diffusion baseline at matched entropy and sampling budget. Reinforce-DMD improves the frontier at larger budgets, reaching a generative perplexity of 14.9 at an entropy of 5.00 nats with 256 NFEs, a 20% reduction under the same comparison protocol.
Sep 30, 2026

PassGPT+: Leveraging Linguistic Priors for Password Modeling

Passwords remain the dominant online authentication mechanism, and understanding how humans choose them is essential for defensive strength estimation and attack simulation alike. Recent learning-based approaches such as PassGAN and PassGPT have shown that deep generative models can learn password structure directly from leaked corpora. However, both train from random initialization on password data alone. The role of linguistic prior knowledge in password modeling, and what it reveals about how humans create secrets, remains largely underexplored. Here, we address this gap with PassGPT+, which adapts the linguistic prior of GPT-2 to password observations through character-aware tokenization. We also introduce PassDiffusion, the first absorbing-state discrete diffusion model for password generation, as a probe of whether non-autoregressive approaches are competitive. On the RockYou benchmark, PassGPT+ recovers 22.53% of held-out passwords at 108 guesses, a 16% relative gain over PassGPT, and retains 79% of this match rate when transferred without retraining to a disjoint 2020 leak dataset, demonstrating that linguistic priors capture persistent regularities of human password generation. PassDiffusion underperforms by two to three orders of magnitude, indicating that autoregressive modeling is substantially better matched than iterative denoising to the discrete, exact-match nature of password generation.
Sep 30, 2026

GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed

Autoregressive (AR) grounding models serialize spatial predictions, introducing sequential latency and imposing a causal order on output tokens. We view grounding as visual evidence extraction: objects, locations, and spatial relations are jointly constrained by the image and query, yet their dependencies do not imply an intrinsic left-to-right generation order. This distinction makes bidirectional diffusion a natural fit, allowing spatial hypotheses to emerge in parallel and be jointly refined through iterative denoising. We introduce GroundAnything, a 4B-parameter grounding foundation model that reconciles fast parallel decoding with precise localization through blockwise denoising. Training combines grounding pretraining from public datasets and dedicated data engines, direct AR-to-diffusion conversion with joint AR and diffusion objectives, supervised fine-tuning, and GRPO-based reinforcement post-training. Across 30 grounding benchmarks, our autoregressive variant, GroundAnything-VLM, establishes a new overall state of the art among similarly sized models at 72.42%, remaining competitive with GPT-6 Astra (71.35%). With entropy-guided decoding, GroundAnything also surpasses the prior state of the art at this scale, averaging 61.75% versus 53.32% for the fast MTP-based LocateAnything model. We further explore decoding strategies, showing that an optional self-speculative mode achieves a 4.51×4.51\times speedup over the AR counterpart with a 0.74 percentage-point drop in COCO F1mIoU. Infrastructure experiments show that progressive inference optimizations translate parallel decoding into practical speedups. These support efficient visual grounding in latency-sensitive real-world systems.
Sep 30, 2026

Self-Repulsive Sampling for Diffusion Language Models

Sampling several responses and voting over their answers can improve a language model's accuracy, but repeated answers limit the benefit of additional samples. Raising temperature increases diversity at a potential cost to per-sample accuracy. We introduce Self-Repulsion (SR), a sampler for masked diffusion language models that uses peer commitments to diversify the pool. At each penalized denoising step, each path lowers a token's logit according to how many peers have committed that token at the same position. Paths share a batched forward pass and then commit in sequence, so later paths observe choices made earlier in the same step. This coupling requires no training or additional forward or backward pass and can produce distinct paths even at temperature zero. When all paths commit a position together from identical logits, the update exactly maximizes total logit minus a convex duplication cost. On LLaDA-8B-Instruct with ten paths and 128 denoising steps, deterministic SR reaches 80.38% plurality accuracy on GSM8K, compared with 70.17% for the unpenalized greedy decoder. At temperature 0.6 and matched model-evaluation budgets, the count penalty improves over self-consistency by 2.06 percentage points in blocks of 32 and 14.50 under pure diffusion. Experiments on GSM8K, MATH and TruthfulQA show that voting gains arise mainly from higher coverage of correct answers, with gains that vary by benchmark and decoding regime.
Sep 29, 2026

Does This Action Still Explain the Task? Reverse Scoring for Diffusion Language Model Agents

Diffusion-based large language models (dLLMs) promise to break the sequential latency bottleneck of autoregressive agents through parallel decoding, but recent evaluations show this efficiency does not transfer to embodied agentic competence: dLLM-backed agents repeatedly fall into retry loops, re-issuing an action long after it has failed. We give a mechanistic account of this failure and a training-free remedy. We trace the retry loop to the adaptivity of masked decoding: the sampler commits the positions it is most confident about and defers the uncertain ones, and at a failure state the context already offers a confident fill for the deferred decision, i.e. the failed action itself, so the retry is committed without the failure feedback ever being confronted. We model the resulting distortion of the action distribution as a task-blind corruption: contextually salient actions (e.g., the action just taken) receive inflated probability by a factor that depends on the state and the action but not on the task. Under this model, we analyse an invariance proposition: the task-blind factor cancels exactly from the reverse conditional, i.e. the likelihood of the task given the state and a candidate action, which coincides with the task posterior of an idealized uncorrupted model. Masked dLLMs evaluate the reverse conditional natively, unlike autoregressive models, by masking the task tokens and denoising, at the cost of a few parallel passes per candidate. We instantiate the rule as Reflect Reverse and evaluate it on four multi-turn embodied benchmarks, where it improves task success and progression rates over forward-scoring baselines.
Sep 29, 2026

DEdit: Iterative Draft Editing for Speculative Decoding

Speculative decoding accelerates autoregressive LLMs by having a lightweight drafter propose tokens that the target model verifies in parallel. Diffusion-based drafters further reduce drafting latency by proposing multiple tokens at once. However, these tokens are predicted independently, so a single early error causes prefix verification to discard the rest of the draft, even when it contains useful downstream predictions. We introduce DEdit, a diffusion-based drafter that can not only draft by conventional parallel unmasking but also iteratively edit its draft through token-to-token predictions. Through editing, later predictions can serve as bidirectional context for repairing earlier errors and extending the accepted prefix. To teach the model to repair errors while preserving correct predictions, we propose ProposalMix, a training scheme that mixes draft predictions with ground-truth tokens based on first-pass confidence during training. Across seven benchmarks on Qwen3-4B and Qwen3-8B, DEdit achieves the highest macro-average token acceptance and speedup among the evaluated drafters, reaching macro-average speedups of 5.72×5.72\times and 5.97×5.97\times over autoregressive generation under greedy decoding, respectively. Further analysis shows that acceptance improves with more editing passes and wider drafting windows, and that ProposalMix halves harmful edits that shorten the accepted prefix. Moreover, restricting the editor to causal attention lowers acceptance, especially on highly predictable outputs, indicating that future context is a key source of these gains.
Sep 29, 2026

Fine-Tuning Diffusion Language Models with Context Selection and Target Weighting

Supervised fine-tuning of discrete diffusion language models masks some response tokens and trains the model to recover their original values from the visible context. The masking pattern therefore determines both the context available to the model and the tokens it learns to predict. Uniform random masking does not explicitly account for the interaction between these choices. We introduce GoldiMask, which selects tokens to reveal as context by approximately maximizing a submodular objective. This objective uses model signals to balance the benefit of revealing tokens against their value as prediction targets. GoldiMask then weights the remaining targets according to how they benefit from the selected context and their remaining learning potential. Across three backbones and three training datasets, GoldiMask achieves the highest average accuracy in most evaluated settings, demonstrating gains on both reasoning and code generation. Component ablations show that both context selection and target weighting contribute to the gains. GoldiMask also reduces decoding iterations on GSM8K and MATH-500 under confidence-threshold parallel decoding, while maintaining comparable accuracy at higher confidence thresholds.
Sep 29, 2026

Acceleration of Diffusion Language Model through Discrete Average Generator

Discrete diffusion models and flow matching have emerged as powerful frameworks for generative modeling over discrete state spaces, yet efficient few-step generation remains a fundamental challenge. In this work, we introduce the Discrete Average Generator, a principled extension of MeanFlow to Continuous-Time Markov Chains (CTMCs). Analogously to how MeanFlow defines an average velocity field over a time interval in continuous spaces, we define an average generator as the normalized increment of the transition kernel over a time interval. We show that this average generator satisfies a self-consistency identity, which provides the foundation for our training objective. We further develop training strategies that align with the standard training paradigm of diffusion language models while keeping the resulting objective tractable. When projected onto per-coordinate marginals, the self-consistency identity admits a closed-form expression, enabling efficient training and inference. In Potts model simulations, our objective reduces the total variation distance of the KK-step sampler by up to 67%. On OpenWebText, our method achieves the lowest generative perplexity among the evaluated methods for 8 to 64 sampling steps while enabling a 16×16\times acceleration, and achieves comparable performance to existing methods on ImageNet.
Sep 29, 2026

Alpha Diffusion Language Models: Factorization Alone Is Not the Problem

Discrete diffusion language models can generate multiple tokens in parallel, but reducing the number of denoising steps can lead to inconsistent predictions. Standard cross-entropy training fits conditional token marginals, whereas parallel generation requires consistent joint predictions. We introduce Alpha Diffusion Language Models (AlphaDLM), trained with a sequence-level alpha loss that recovers cross-entropy in the limit of vanishing alpha and has a joint-mode optimum at alpha one. Our analysis characterizes how the objective and factorization jointly determine the fitted distribution. We identify conditions under which intermediate alpha preserves multiple valid completions while excluding invalid token combinations. Trained on TinyGSM, our method achieves 34.6% accuracy on GSM8K with only four model evaluations. We further scale the method to SDAR-1.7B and evaluate it on code and mathematics benchmarks. These results show that changing the training objective can improve the accuracy-computation trade-off of factorized diffusion language models.
Sep 29, 2026

On Trajectory-Aware Training for Masked Diffusion Language Models

Masked diffusion models (MDMs) generate text by unmasking several tokens per step, but they are trained and sampled under different conditions. The model is trained on randomly masked sequences, whereas inference follows a trajectory shaped by the model's own predictions. Additionally, each step has no access to what the previous one computed. Recent methods narrow these limitations from separate angles, leaving open how these choices interact. We introduce PUMBA, a unified framework for trajectory-aware training that trains the denoiser on consecutive steps of policy-induced trajectories, passes information between steps, and optimizes them jointly by backpropagation through time. A controlled study of this design space shows that i) exact train--inference alignment fails due to local overfitting, whereas a looser alignment still brings training masks closer to those seen at inference; ii) passing continuous information outperforms discrete gradient estimators through the commitment at each step; and iii) performance improves as backpropagation through time spans more steps, which we support theoretically. Combined, these components match the best checkpoint of a same-size autoregressive model. Building on these findings, we scale PUMBA to supervised fine-tuning of LLaDA-8B, where it improves the trade-off between performance and number of function evaluations (NFEs) in both full-canvas and block diffusion generation. At matched performance, it needs up to 22% fewer NFEs than standard fine-tuning with twice the budget in full-canvas generation, and up to 26% fewer than standard fine-tuning for the same number of steps in block diffusion.
Sep 29, 2026

Time-Anchored Diffusion Language Models: Latent-Space Caching for Fast Generation

Recent work on anchored diffusion language models improves denoising by shaping an intermediate latent space with supervised important-token targets. In this work, we introduce time-based (self-supervised) anchoring, which learns and reuses latent anchors without requiring such targets. Our key observation is that anchors encode persistent properties of the clean sequence, such as its semantic intent, global structure, or intermediate plan. Although their hidden representations become stale as the token canvas evolves, their semantic content remains useful across nearby diffusion times. This is implemented through a two-stage architecture consisting of a relatively expensive anchor network that generates the latent cache state and a lightweight denoising network that intelligently combines the cached latent state with the current state at each reverse step using a fusion module. This gives anchoring a latent-space caching interpretation: the anchor network is evaluated periodically, while its cached representation is reused across multiple reverse steps. We instantiate this framework as TADM:Post-train, which time-anchorizes pretrained DLMs, and TADM:Pretraining, which learns time-based anchors during pretraining. Applied to DiffusionGemma-26B, TADM:Post-train improves throughput by approximately 49% to 79% on several math, code, and STEM benchmarks (GSM8K, AIME26, GPQA-Diamond, LiveCodeBench-v6, HumanEval, MMLU-Pro). TADM:Pretraining reduces Transformer-layer computation by up to 38% relative to a standard single-stage DLM, achieves up to 73% higher measured throughput than ADLM.
Sep 29, 2026

E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models

Masked diffusion models (MDMs) generate sequences by progressively unmasking several tokens per denoising step, but their reverse process is typically factorized over positions, limiting sample quality in the few-step regime where diffusion's speed advantage over autoregressive decoding matters most. A recent line of work introduces a continuous Gaussian latent, trained as a variational autoencoder, to capture correlations across positions, but such approaches are prone to posterior collapse, where the latent is silently ignored. We propose Enhanced Mixture-of-Experts (E-MoE), which builds the reverse process as a mixture of factorized distributions over a discrete shared latent given by the expert-routing decisions of a Mixture-of-Experts (MoE) backbone, without increasing active parameters over the factorized baseline. Across synthetic multi-modal benchmarks, binarized MNIST, and LM1B, E-MoE improves few-step generation over factorized baselines.
Sep 29, 2026

Rethinking Soft Tokens for Parallel Decoding in Diffusion Language Models

Diffusion language models (DLMs) enable parallel generation by predicting and committing multiple tokens at each denoising step, yet they can generate individually plausible but mutually inconsistent tokens. Recent work shows that soft tokens can mitigate this issue by representing uncertain positions with continuous embeddings built from the model's predictive distribution at the previous decoding step. However, although soft tokens are commonly understood as preserving predictive uncertainty, how soft-token feedback improves parallel decoding has not been systematically examined. In this paper, we investigate this question in frozen pretrained DLMs to examine soft-token feedback without the effects of additional training. To construct soft-token inputs in a training-free setting, we identify a geometric mismatch between conventional soft-token construction and the pretrained embedding space. Based on this observation, we propose a training-free, geometry-aware construction of soft tokens. Our analysis of soft-token feedback suggests that uncertainty preservation alone does not fully explain how it reshapes subsequent predictions. To better explain how soft-token feedback improves parallel decoding, we provide empirical evidence that it favors coherent token sequences. Across four pretrained DLMs and four math and code benchmarks, our method outperforms standard parallel decoding and a training-free Euclidean soft-token baseline. Code: https://github.com/kodaikawamura/rethinking-soft-tokens
Sep 28, 2026

HyperZip: Efficient Data Compression through Personalized Diffusion LLMs with Hypernetworks

Large language models (LLMs) have shown strong potential for lossless data compression, but existing approaches are constrained by the high computational cost and low throughput of autoregressive decoding. We propose HyperZip, an efficient and scalable LLM-based compression framework that leverages diffusion-based LLMs (dLLMs) with Multi-Token Prediction (MTP) to accelerate LLM-based data compression processes. We identify a trade-off in diffusion-based compression, where increasing decoding throughput degrades the compression rate. To mitigate this trade-off, HyperZip employs a hypernetwork to generate data-specific updates from a context representation, adapting the dLLM to the target data without costly fine-tuning, resulting in a low compression rate and high throughput. Extensive experiments show that HyperZip achieves a superior trade-off between compression rate and speed compared with state-of-the-art baselines.
Sep 28, 2026

Reasoning with Continuous Latent Diffusion

We introduce the Continuous Embedding Diffusion Reasoner (CEDR), a training and inference pipeline for continuous latent diffusion models that achieves strong performance on mathematical reasoning and code generation. We learn compact answer representations from three hidden layers of the autoregressive Qwen3-4B model. A staged curriculum first trains the ELF-based diffusion backbone with teacher prompt conditioning, then initializes a compact prompt encoder through imitation and jointly adapts both networks. This replaces the large teacher contextual prompt encoder at inference with a compact learned Transformer. The multilayer answer representation also enables asynchronous inference, with different components denoised under separate clocks. We further adapt DiffusionNFT to learned self-conditioning guidance, enabling reinforcement learning with verifiable rewards. Controlled ablations evaluate the effects of representation design, prompt training, and denoising schedules. CEDR outperforms recent continuous-diffusion baselines at comparable backbone scales, including before reinforcement learning. With a 638M-parameter denoising backbone and DMD2-style stochastic sampling, post-NFT CEDR-L achieves {pass@1, pass@8} accuracies of {68.97%, 80.82%} on GSM8K and {32.40%, 53.80%} on MATH500 at 64 denoising steps per sample, and {37.56%, 51.00%} on HumanEval+ and {26.96%, 39.92%} on MBPP+ at 128 steps per sample. Code and model weights are available at https://github.com/chengxiang/CEDR and https://huggingface.co/xc91/CEDR, respectively.
Sep 28, 2026

FluxLite: Inference-Time Proposal Control for Discrete Diffusion Models

Many inference-time tasks for pretrained discrete diffusion models and diffusion language models reduce to drawing samples from a tilted version of the pretrained distribution. Feynman-Kac sequential Monte Carlo (SMC) makes this correction exact in principle, but its prescribed weights routinely degenerate when the proposal dynamics are misaligned with the tilt, capping the practical gains from additional particles. We introduce FluxLite, a lightweight, training-free proposal-control framework for discrete diffusion. On the sparse directed graph of pretrained reverse rates, any sparse jump-rate perturbation can be exactly compensated by a qtq_t-weighted graph-divergence term in the Feynman-Kac potential; the target path is therefore preserved while the residual reweighting variance becomes a local convex objective. We instantiate this principle as two practical samplers: a one-hop local reallocation rule (HEU) and a small nonnegative quadratic program over pretrained-rate bases (D-VCG). We further prove population stability under the standard score-entropy training loss, identifying a tilted-path coverage factor that governs robustness to score error, together with finite-particle convergence for a fixed controlled Feynman-Kac recursion. Empirically, FluxLite improves over standard Feynman-Kac SMC baselines by up to two orders of magnitude in terminal KL on an analytically tractable finite-state CTMC benchmark, and reduces row-correlation MSE on 2D Ising sampling by 5-7x in geometric mean and up to 55x at peak.
Sep 28, 2026

Simplex Diffusion Models

Diffusion models have revolutionized generative modeling for continuous data through the gradual refinement of a belief state. This iterative refinement has not yet carried over to discrete diffusion models, which discard uncertainty at intermediate steps through categorical sampling (information collapse). We propose Simplex Diffusion Models (SDMs), a framework that lifts the diffusion process to the probability simplex to represent beliefs over categories. SDMs admit probability paths with closed-form reverse transitions and can be trained with a simple cross-entropy loss. Contrary to earlier proposals such as Dirichlet Flow Matching which requires integrating an ordinary differential equation, we introduce a DDIM-like sampler with a tunable level of stochasticity. Because SDMs operate on samples on the simplex, they can carry uncertainty across denoising steps, which mitigates information collapse. On OpenWebText, SDMs are competitive with strong Discrete Diffusion baselines, achieving 17.017.0 GenPPL at 5.465.46 unigram entropy in 64 sampling steps, close to real validation data. Even without Self-Conditioning (SC), SDMs outperform masked and uniform diffusion (with SC or predictor-corrector sampling) on code generation (TinyGSM, T=0.1T=0.1; 49.0%49.0\% vs. 45.8%45.8\%). Distilled down to 8 steps, SDMs solve 32.1%32.1\% of GSM8K problems, more than distilled Discrete Diffusion models with 128 steps (21.4%21.4\%).
Sep 28, 2026

Why Deterministic PRM Guidance Underperforms in Discrete Diffusion Reasoning

Discrete diffusion language models (dLLMs) expose a denoised solution at every step, which makes process reward model (PRM) guidance look like a way to spend compute at test time. We show that once denoising, PRM scoring, and outcome reward model (ORM) scoring are charged in the same budget of forward passes, its deterministic form loses to a much simpler baseline. Our PRMs score intermediate denoising states and are trained on the correctness of the final answer. On Dream-v0-Instruct-7B with 8 candidates per GSM8K problem, keeping the candidate with the highest PRM score at every scoring step reaches 65.18%, while independent sampling plus an ORM reranker trained for the task reaches 75.13%. The gap grows to 12.69 percentage points (pp) with 32 candidates, and is 9.85 pp on MATH and 12.16 pp on MBPP. We trace it to two separable failures. First, guidance prunes on a weak signal: on GSM8K, PRM ROC-AUC falls from 0.77 to 0.54 as the mask ratio rises, a decay that persists when states are relabeled with fresh rollouts, and pruning lowers the best accuracy reachable from the candidate pool from 81.05% for independent samples to 67.30%. Second, on GSM8K and MATH, the PRM is a poor final judge: a sequential Monte Carlo sampler at the same budget restores that ceiling to 77.89%, yet selecting with the PRM gives 65.48%, on par with deterministic guidance, while a PRM retrained on final states matches the ORM on identical candidates. MBPP separates the two: there the PRM reaches 65.47% when reranking finished programs, on par with the ORM, but 50.88% when it guides denoising. The results point to two targets for dLLM guidance: keep correct partial solutions alive through early denoising, and leave the final choice to a verifier trained on final states. We release the corpus of denoising states with outcome labels and evaluation toolkit for reproducible comparisons at matched compute.
Sep 28, 2026

d-OPD: Future-Aware On-Policy Distillation for Block Diffusion Language Models

Large language models (LLMs) typically generate text autoregressively (AR), predicting one token at a time. Block diffusion language models (dLLMs) instead generate blocks sequentially while denoising multiple tokens in parallel within each block, offering a promising way to accelerate generation. Rather than training such models from scratch, recent work adapts strong pretrained AR models into block dLLMs through distillation. On-policy distillation (OPD) has been widely used for LLM training because it supervises the student on states generated by its current policy, rather than only on fixed offline trajectories. By training on the states the student actually visits, it reduces the mismatch between training and generation and can provide more relevant supervision as the student evolves. Recent work has extended this idea to AR-to-block-diffusion conversion. However, this setting introduces a fundamental mismatch in supervision: the block-diffusion student and the causal AR teacher condition on different information at the same training state. The student predicts from the entire partially denoised block, including visible future context, whereas the standard AR teacher target is defined only from the causal prefix. As a result, the teacher distribution used for distillation is not fully aligned with the information available to the student. We therefore introduce d-OPD, a future-aware on-policy distillation method that corrects the AR teacher distribution to better align with the student-visible state by incorporating visible future information within each block, providing supervision that better matches the information used by the student. Across Qwen3 models from 0.6B to 8B, d-OPD improves the six-benchmark average by up to 4.04.0 points over OPDLM and reduces training time by 1.351.35-1.58×1.58\times. The code is available at https://github.com/mit-han-lab/d-OPD.
Sep 28, 2026

Learning to Re-Draft: A Variational Stackelberg Game for Discrete Diffusion

Discrete diffusion models offer the ability to re-draft, revisiting and correcting earlier tokens throughout generation. This capability depends on the forward corruption process that defines what the denoiser learns to correct. Masked diffusion models fix tokens once they are unmasked, while uniform diffusion permits revisions but relies on uniformly random token substitutions. We instead learn which substitutions are most useful for training the denoiser to re-draft. We introduce Variational Stackelberg Discrete Diffusion (VSDD), a framework for learning a semantically aware corruption process. VSDD formulates training as a leader-follower game: the leader defines a Markovian corruption process parameterized by the denoiser's token embeddings, while the follower optimizes a variational denoising objective with the corruption process held fixed. The leader rewards corruptions based on how much the denoiser improves after learning from them, rather than on how easily the current denoiser can reconstruct them. We measure this improvement under a fixed reference corruption process, approximate the follower's response with a one-step gradient update, and optimize the leader using a score-function estimator. We evaluate VSDD across molecular, text, and playlist generation. VSDD substantially improves molecular validity over uniform and masked diffusion, reduces text perplexity relative to uniform diffusion while remaining competitive with masked diffusion, and achieves sizable improvements in offline playlist recommendation metrics.
Sep 28, 2026

BV Loss: Block Verification-Aware Loss for Block Diffusion Speculative Decoding

Diffusion drafters accelerate speculative decoding by proposing multiple tokens in parallel. Despite recent advances in speculative decoding through sequence-level drafting and verification, existing training objectives remain largely designed around token-level verification. To address this mismatch, we introduce Block Verification-aware loss (BV loss), a training objective designed to maximize the expected acceptance length of a drafted sequence. BV loss is directly derived from the block verification acceptance rule, providing a principled connection between the drafter training objective and the inference-time verification mechanism at the sequence level. Across math, code, and chat benchmarks, BV loss increases the mean number of tokens accepted per verification call under block verification by 13.0--21.0% over cross-entropy loss training for DFlash and DSpark with Qwen3-4B and Qwen3-8B without changing the inference procedure. BV loss also outperforms tokenwise acceptance objectives such as TV loss and LK loss, and its gains extend to token verification and greedy decoding. These results demonstrate the benefit of training block diffusion drafters with an objective aligned with sequence-level verification, rather than optimizing each token independently.
Sep 28, 2026

Low-Confidence Remasking Traps Flexibility: Realizing Arbitrary-Order Potential for Diverse Rollouts in Diffusion LLMs

Masked diffusion language models support arbitrary-order generation, suggesting a natural way to produce diverse outputs. However, recent work argues that this flexibility reduces diversity by delaying high-uncertainty tokens that can lead to different generation paths. We trace this diversity loss not to arbitrary-order generation itself, but largely to low-confidence remasking (LCR), a widely used decoding rule. At each step, LCR samples a token at every masked position but commits only the sampled token with the highest probability, filtering out the rest. We show that this mechanism can exponentially suppress lower-probability tokens as more positions compete, and observe the same suppression in LLaDA. In contrast, top-probability position selection (TPP), which has often been conflated with LCR under the shared label confidence-based decoding, avoids this diversity loss. TPP first selects the position whose most likely token has the highest probability, then samples directly from that position's distribution. Replacing LCR with TPP restores diversity and yields Pass@kk comparable to left-to-right decoding, suggesting that the reported diversity loss stems largely from LCR's filtering rather than from generating high-confidence positions first. To further exploit order flexibility, we introduce Entropy-Guided Initialization (EGI), which samples the first token at the highest-entropy position and then follows TPP. This simple modification further improves rollout diversity and solution coverage beyond left-to-right decoding, with gains extending to downstream policy optimization, highlighting the potential of arbitrary-order generation for diverse rollouts.