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.

Nov 12, 2026

Learning When Diffusion Language Models Should Regenerate

Remasking allows diffusion language models to revisit generated text, but effective correction requires knowing when and what to revise. Our diagnosis on Sudoku reveals two gaps. Hidden states contain error information that token confidence does not reliably expose, and remasking an isolated error usually reproduces it when the surrounding context remains fixed. These findings motivate DiVR , a verifier that predicts whether a partially generated sequence will reach a correct answer. Trained on the final correctness of the model's own trajectories, DiVR decides when to regenerate a block and which alternative to retain, while the generator itself stays unchanged. On five reasoning benchmarks with LLaDA-8B-Base, DiVR raises the average accuracy of the fine-tuned generator from 35.3% to 43.4% at 1.91.9-3.5×3.5\times the cost of greedy decoding, above search baselines that use 3.83.8-7.4×7.4\times. It also complements post-training such as supervised fine-tuning and reinforcement learning, and generalizes to other generators, including Dream.
Nov 11, 2026

Schedule-Calibrated Commitment Order in Blockwise Masked Diffusion Language Models

Sequential block schedules can make token commitment in masked diffusion language models appear strongly left-to-right even when within-block order is random or reversed. For equal contiguous blocks processed without interleaving in positional order, with untied irreversible commitments, we use a Spearman-rank identity to express global position-commitment correlation as a deterministic schedule reference plus a weighted mean within-block correlation. In a reported paired study of LLaDA-8B-Instruct on 500 math and code prompts with 256 generated positions and updates, increasing block size from 64 to 128 lowers mean global correlation from 0.9845 to 0.9149. Uniform-random within-block references are already 0.9375 and 0.7500. After calibration, mean within-block correlations are 0.751 and 0.660, so the larger raw residual at block size 128 does not indicate stronger local positional ordering. We give sharp bounds and derive calibrated intervals from reported summaries, while documenting the limitations of the aggregate-only empirical release. The analysis supports schedule-aware interpretation of commitment order; its empirical scope is one checkpoint and two schedules, without evaluation of output quality or inference speed.
Nov 11, 2026

Mask-Pattern Position Augmentation: Training-Free Long-Context Extrapolation for Diffusion LLMs

Diffusion large language models (dLLMs) generate by iteratively unmasking tokens with bidirectional attention, but rotary position embeddings (RoPE) still confine them to a fixed pretraining window. Existing training-free extrapolation, notably LongLLaDA, copies NTK-based RoPE scaling from autoregressive (AR) models and applies one static, length-only transform at every position and denoising step. That transplant ignores a signal unique to dLLMs: the evolving mask pattern of the partially decoded sequence. We propose Mask-Pattern Position Augmentation (MPPA). MPPA linearly interpolates indices into the pretrained window, then adds a per-position, per-step offset from cumulative mask density. The offset restores rotary phase resolution that interpolation collapses, and it shrinks as tokens are committed, so late steps see near-clean interpolated geometry. On LLaDA-8B-Instruct at twice its 4096-token window, MPPA reduces perplexity from 19.51 under standard decoding to 3.12 at 8192 tokens, below the in-window figure, with 100% passkey retrieval and unchanged needle-in-a-haystack accuracy. These preliminary 2×2\times results motivate mask-conditioned position engineering as a diffusion-native path to long-context generation.
Nov 7, 2026

Global Context Guidance for Diffusion Large Language Models

Diffusion large language models (LLMs) have recently emerged as a promising alternative to autoregressive LLMs, enabling parallel generation through iterative denoising. Recent guidance methods improve diffusion LLMs by introducing a weak model to guide the prediction of the base model during denoising. However, the influence of the weak model during denoising remains less clear, and computing its prediction requires an additional forward pass at each denoising step. We present in this paper an in-depth analysis of how diffusion LLMs use contextual information during denoising. Our analysis reveals that diffusion LLMs rely more heavily on local context while often underutilizing global context. Based on this observation, we introduce Global Context Guidance (GCG), a feature-space guidance method that enhances global-context utilization without constructing an explicit weak model. Extensive experiments on LLaDA-8B, Dream-7B, and LLaDA-1.5 across diverse benchmarks demonstrate that GCG consistently outperforms existing guidance methods, while incurring much lower inference overhead.
Nov 7, 2026

Test-Time Scaling for Diffusion Language Models with Frequency-Aware Remasking

Diffusion language models (DLMs) generate text through iterative denoising, where decisions about which tokens to commit or remask strongly affect output quality. Existing certainty-aware strategies rely on instantaneous confidence or entropy, but these local signals can be misleading on difficult reasoning problems. We formulate DLM test-time scaling as adaptive allocation of both inference compute and decoding strategies. Our two-stage approach uses a fast base sampler to solve easy questions, then applies frequency-aware remasking to the remaining hard questions. This strategy tracks how often each token's prediction changes across denoising steps and preferentially commits frequently changing tokens, encouraging samples to explore alternative reasoning paths rather than collapse onto the model's preferred trajectory. Experiments with LLaDA-8B-Instruct and Dream-7B-Instruct show consistent improvements over a strong adaptive allocation baseline, with absolute accuracy gains of up to 3.73 percentage points (LLaDA) and 7.33 percentage points (Dream) on MATH-500, and absolute coverage gains of up to 11.39 percentage points (LLaDA) and 10.98 percentage points (Dream) on HumanEval. On questions unsolved by the base certainty-aware strategy after 100 samples, frequency-aware remasking recovers solutions for 3.75% to 18.75% of examples.
Nov 5, 2026

SpecCTMC: Exact Speculative Simulation of Diffusion Language Model Continuous-Time Markov Chains

Masked diffusion language models offer parallel text generation, but independently sampling multiple tokens can disrupt their dependencies and change the output distribution. Exact continuous-time sampling avoids this approximation, yet typically requires one sequential model evaluation per token. We introduce SpecCTMC, a training-free speculative sampler that reduces this sequential bottleneck without a separate draft model. SpecCTMC proposes successive token updates using previously computed model predictions, evaluates their intermediate states in parallel, and applies an exact acceptance-and-correction procedure. We prove that SpecCTMC preserves the joint distribution of event times, generation order, and token values of the underlying continuous-time Markov chain. It therefore introduces neither time-discretization error nor additional factorization error from parallel decoding. Experiments on LLaDA-8B-Instruct and Dream-v0-Instruct-7B across five benchmarks demonstrate 1.73-4.38x fewer sequential model evaluations than direct continuous-time sampling, while preserving the reference sampling distribution. Our code is available at: https://github.com/kasakh/specctmc
Nov 5, 2026

Masked diffusion as random sequential adsorption: an exactly solvable model for parallel decoding

Masked diffusion models generate text by reversing a stochastic masking process. Unlike autoregressive LLMs, which generate text one token at a time, they can reveal many tokens in parallel, at the price of an error, since tokens revealed together are drawn independently. We show that this trade-off maps onto the statistical physics of driven many-particle systems. (i) For any data, the error of parallel generation is the entropy production of the generation process relative to the reversed masking process. (ii) For Markov data, generation is a random sequential adsorption process with labelled gaps, which we solve exactly for any number of steps. (iii) For many steps, the error becomes the dissipation of a slowly driven system: the unmasking schedule plays the role of the driving protocol, the number of steps that of its duration, and the error is set by a thermodynamic metric that we obtain in closed form from the mutual information between tokens. Optimal schedules are geodesics of this metric. The exact solution shows how correlations in the text shape both the error and the quality of the output. The longer the correlations, the more the error concentrates at the start of generation. Schedules that start with small steps then gain more, both in path error and in perplexity, and need fewer steps to reach their asymptotic behaviour.
Nov 4, 2026

Training-Free Character-Length Control in Summarization with Diffusion Language Models

Summarization with a fixed character length is important for tasks such as headline and article preview generation, where text must fit an exact space. Autoregressive language models (LMs) struggle with precise character-length control: training-based methods require task-specific fine-tuning, while zero-shot prompting strategies achieve only approximate compliance. We propose DELTA (Diffusion Expected Length Targeting Algorithm), a training-free inference-time method for masked diffusion LMs that estimates expected character length at each denoising step and dynamically inserts or removes masked positions to steer toward a target length. DELTA requires no fine-tuning and is designed for off-the-shelf masked diffusion LMs. Across three summarization benchmarks, it reduces character-length errors by an order of magnitude compared to baselines while maintaining competitive summary quality in both automatic and human evaluation.
Nov 3, 2026

VIGOR: Visual Gain Ordering for Hallucination Mitigation in Multimodal Discrete Diffusion Language Models

Multimodal discrete diffusion language models (dLLMs) generate responses through iterative mask prediction and offer a promising alternative to autoregressive large vision-language models. However, they still suffer from hallucination, producing outputs that are plausible under language priors but insufficiently grounded in the image. We argue that hallucination in multimodal dLLMs is not only a token prediction problem, but also a premature commitment problem: during iterative denoising, some masked positions are unmasked too early because their confidence is driven more by language priors than by visual evidence, and these early commitments can bias later generation. To address this issue, we propose Visual Gain Ordering (VIGOR), a training-free and plug-and-play inference strategy that reorders masked positions according to counterfactual visual evidence. At each denoising step, VIGOR compares the confidence of each candidate token under the original image-conditioned pass and an additional visual-ablation pass, and prioritizes positions whose predictions are more strongly supported by the image. Experiments on LLaDA-V and Lumina-DiMOO show that VIGOR consistently reduces hallucination and improves multimodal reasoning performance across benchmarks. Code is available at https://github.com/THU-BPM/Vigor.
Nov 3, 2026

Bidirectional Centered Score Matching for Discrete Diffusion Distillation

Discrete diffusion models with parallel decoding are often distilled using token-wise KL objectives. Although token-wise KL exactly matches the corresponding product-form block distributions, its probability-space optimization geometry and teacher-only candidate truncation may provide insufficient supervision under aggressive few-step distillation. Concrete Score Matching (CSM) provides an alternative within-position objective by aligning relative logit geometry, but its complete pairwise form incurs \mathcal{O}(|\mathcal{V}|^2) complexity. We propose Bidirectional Centered Score Matching (BCSM), an efficient within-position distillation objective that constructs an \mathcal{O}(K) bidirectional active set, reduces pairwise comparisons to \mathcal{O}(K^2), and factorizes them into an \mathcal{O}(K) objective. We further introduce cross-timestep transition alignment. Experiments demonstrate strong quality--efficiency trade-offs, including up to 6\times NFE reduction with a 16.9% relative HumanEval gain and a 15.1% relative MMMU gain at single-step inference.
Nov 2, 2026

Train the Sampler You Run: Few-Step Masked Diffusion Language Models via Mixture-Likelihood Fine-Tuning

When a masked diffusion language model is sampled in a few parallel steps in a random order, the model actually run is a mixture over reveal plans, while cross-entropy training fits token conditionals rather than this mixture. We introduce mixture-likelihood fine-tuning (Mix-FT), which fine-tunes a pretrained model on an importance-weighted bound of its own sampler's likelihood, with no teacher and no change to the architecture or the sampler. On LLaDA-8B, a LoRA fine-tune for 0.002% of the pretraining compute writes better OpenWebText continuations than continued cross-entropy and Alpha diffusion given the same compute, at every tested step count. With half the model evaluations, 16 rather than 32 steps per block, its continuations score better than the original model's in every seed under three judges, with confidence-based decoding as well as random order, and under confidence-based decoding two blinded LLM raters from different model families prefer them in 69 and 71% of pairs. On a public 110M-parameter block diffusion model, Mix-FT takes under four T4-hours and beats continued cross-entropy, Alpha diffusion and SDTT distillation at 4, 8 and 16 steps per block, under two judges in every seed. Exact likelihoods on text8, with a control that weights every plan equally, attribute about three quarters of the likelihood gain over continued cross-entropy to weighting each plan by how well it explains the sequence.
Oct 31, 2026

Step-dLLM: Adaptive Step-aware Sparse Attention for Efficient Diffusion LLM Inference

Diffusion large language models (dLLMs) generate tokens in parallel via iterative denoising, but their bidirectional attention prevents the KV caching used by autoregressive models and recomputes the full attention map at every step, incurring quadratic per-step cost. Existing sparse attention methods for dLLMs construct a sparse mask after an initial full-attention warm-up and reuse it for the remaining steps, while allocating the sparsity budget uniformly across heads. We argue both choices ignore key structural properties of attention in dLLMs: empirically, attention patterns are only locally consistent across steps and shift abruptly at certain transitions, and different heads concentrate their attention mass to very different extents. To this end, we design Step-dLLM, which periodically refreshes a block-sparse mask at anchor steps along the denoising trajectory and reuses it within each locally consistent interval, while employing a head-adaptive global budget that more aggressively compresses heads with concentrated attention. Backed by a fused IO-aware Triton kernel for block-sparse attention, Step-dLLM applies to pre-trained LLaDA-1.5 and Dream-7B-Instruct without retraining, achieving a 2.78x end-to-end speedup over dense FlashAttention on LLaDA-1.5 at 8K context with 256 denoising steps and 90% sparsity, while improving the accuracy-latency trade-off over prior sparse-attention methods.
Oct 30, 2026

BYOD: Build Your Own DLM - Efficient Autoregressive-to-Diffusion Language Model Conversion

Unlike autoregressive (AR) models, diffusion language models (DLMs) permit parallel token updates, flexible generation, and bidirectional attention, but leading systems rely on costly pretraining or full-model fine-tuning. We present Build Your Own DLM (BYOD), a generalizable LoRA-only recipe for parameter-efficient AR-to-DLM conversion. Without architecture-specific tuning, BYOD induces diffusion generation in Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B; to our knowledge, this is the first reported AR-to-DLM conversion of a Mistral model. The adapters comprise 4.1-5.8% of final parameters while every base-model parameter remains frozen; each 25,000-iteration conversion needs at most 102.4M tokens, processed in 3.1-4.0 hours on one Google Colab GPU. All four models generate from fully masked sequences, use bidirectional context in a controlled attention test, and retain substantial parent-model knowledge. BYOD-Gemma is the strongest BYOD conversion on our nine-task evaluation, reaching 50.6% average accuracy compared with 51.9% for LLaDA 8B Instruct in the same harness. Open-generation quality at low network function evaluation (NFE) budgets remains the principal limitation: BYOD perplexity is 8.3-11.1 versus 4.1 for LLaDA and 2.2-3.0 for the AR parents when generating 128 tokens in 64 steps. To enable further experimentation and inspection, we provide a standalone training and evaluation implementation and interactive demos for all four conversions. BYOD thus provides a practical route for extending diffusion generation to other existing model families while identifying the gap that efficient conversion must next close.
Oct 30, 2026

Sparse Text Diffusion under Dual Budgets: Context Connectivity and Expert Activation

Long-context diffusion repeatedly pays for both token interaction and feed-forward computation. We formulate these costs as two separate budgets: the edges available to an attention graph and the experts activated for each token. This view connects local, dilated, and global attention with conditional expert computation without assuming that sparsity alone preserves document-level information. We specify normalized masked attention, distinguish denoising depth from sampling iterations, and derive graph and compute bounds. A reported sparse-expert configuration, denoted DualBudget, is analyzed across scientific summarization, multi-hop question answering, dialogue, and sentence-pair prediction. All supplied table measurements are retained. The length study shows gains over the listed baselines through 16K, together with measurable degradation as length increases. Ablations reveal mixed changes across the three reported scores, while the cost record indicates a 20% reduction relative to DiffuSeq under an unspecified timing convention. The analysis separates these observations from unmeasured claims about training speed, absorbing states, global coherence, and deployment memory.
Oct 29, 2026

Homebrewed SLMs: Autoregressive and Masked Diffusion Language Models Built at Home for Portuguese

Building a language model of one's own no longer requires a data centre. This tutorial trains one from scratch, tokenizer included, on one consumer GPU, and shows how to tell whether it is good for the tasks of a specialist domain. It proposes no new method. It releases a tested Python package and two notebooks in Brazilian Portuguese with autoregressive and masked diffusion generation on one network, a preregistered comparison protocol, legal tasks built from text alone with a Kneser-Ney floor and public yardsticks, reference numbers for models of 11 to 12M parameters, and a catalogue of defects that only measurement revealed. On Brazilian prose and federal laws, at minutes per run, the autoregressive mode peaks at 64 epochs over the same text and then memorises it, while masked diffusion improves up to 256. At their best points on test, the diffusion bound exceeds the autoregressive likelihood by 0.06 bits per byte, within the preregistered minimum effect. On held-out laws and on acts published after training, the model ties a 6-gram floor in likelihood on the same windows, and it loses to the floor at continuing those acts. It beats the floor at detecting tampered text, and fills cross-references more often than the floor and public models of 160M and 630M parameters, only in decrees, a genre absent from training, where those scorers stay near or below chance, and in filling the blank only when each group weighs the same, since per item it ties the 630M model. On laws, the public models, trained with about 5x10^2 to 10^5 times its compute, lead in tampered text and continuation. The blind human reading is preregistered and not part of this version.
Oct 29, 2026

DWISE: Denoising Workload-Informed Scheduling Engine

Unlike autoregressive large language models (AR-LLMs), diffusion LLMs (dLLMs) generate text by iteratively denoising many tokens in parallel, offering a promising path to efficient inference. Existing serving frameworks, however, schedule dLLM requests with policies designed for AR models, namely first-come-first-served (FCFS) admission and prefill priority. We find that both policies are ill-suited to dLLMs, wasting much of their speed. FCFS ignores that a dLLM request's decoding cost is highly heterogeneous and unknown at arrival, causing severe head-of-line blocking. Prefill priority, widely used in AR-LLM serving to keep time-to-first-token low, sharply cuts dLLM efficiency, as rounds that run only prompts leave most of the batch idle. We present DWISE, a Denoising Workload-Informed Scheduling Engine for efficient dLLM serving. DWISE schedules requests by their remaining work, measured in a block-level unit shared by prefill and denoising, so both phases share a round. To predict the unknown denoising workload, DWISE combines a lightweight prompt-based predictor at admission with a single refinement from the first denoising round's hidden states, which capture model-specific information absent from the prompt and require no additional model execution. Implemented in SGLang and evaluated on five models from three dLLM families under bursty workloads, DWISE reduces mean end-to-end latency over SGLang by 5.4×\times on average and up to 14.9×\times while preserving task accuracy. The code will be released publicly.
Oct 27, 2026

Metacognitive Decoding: Efficient Inference for Diffusion Language Agents

Diffusion language models offer a promising path to efficient agents through parallel token generation. However, efficient interaction requires adapting both where computation is spent and how much text is generated. We introduce a metacognitive controller that guides diffusion generation according to the needs of the current response and the agent's interaction with the environment. It learns when to finalize tokens and how far generation should extend, reducing overall computation within responses. Our controller, a lightweight plug-in for the diffusion backbone, is trained on agent trajectories to balance correctness and generation cost. This approach connects the agent's generation process to the interaction the agent serves. Experiments across five benchmarks demonstrate that our approach reduces inference FLOPs by 20.6% on average compared with the most efficient baseline, while maintaining competitive task accuracy. Ablation studies further validate the effectiveness of our approach.
Oct 27, 2026

TDLM: Adaptive Temporal Integration for Diffusion Language Models

Masked diffusion language models repeatedly compute representations for unresolved tokens, but conventional denoisers do not explicitly retain these representations for reuse across steps. We introduce Temporal Self-Conditioning Diffusion Language Model (TDLM), which lets each unresolved position attend to its own history of hidden representations within a frozen denoiser. This gives later predictions access to computation already performed. Training uses self-generated trajectories and a simple cross-entropy objective for tokens revealed between consecutively sampled decoding steps, while causal temporal attention permits parallel evaluation during training. On LLaDA, the temporal modules add only approximately 0.84% trainable parameters. Across LLaDA and Dream, the main comparison shows up to 7.15×\times faster generation than the corresponding original diffusion language model, with competitive answer quality on the evaluated tasks. These results motivate temporal memory as a practical way to make more effective use of repeated denoising computation.
Oct 27, 2026

HiLo-RD: Cross-Budget Self-Distillation for Masked Diffusion Language Models

We propose HiLo-RD (High-to-Low Budget Representation Distillation) to improve masked diffusion language models (MDLMs) under a limited generation budget. Unlike autoregressive models, MDLMs choose both what to generate and where to commit it. This creates two coupled challenges: predicting useful content before reasoning is complete, and deciding when to fix answer tokens that shape later predictions. We identify budget-induced premature commitment: even when a correct solution fits, the model can fix answer fragments near the window boundary before resolving intermediate steps. These fragments become context for later predictions and redirect subsequent reasoning. Existing post-training methods do not directly transfer the stronger trajectories that the same model follows with a larger window. HiLo-RD addresses this gap by (1) pairing large-window teacher states with cropped small-window views and (2) matching their hidden representations at committed positions. This supervision guides both token prediction and commitment order without an external teacher. Across mathematics benchmarks, HiLo-RD improves low-budget accuracy and shortens responses; these gains extend to code generation without code-domain training. On GSM8K, accuracy rises by 9.4 percentage points as mean readable length falls by about 25%. Attention diagnostics and fixed-budget order interventions support the proposed mechanism. Together, these results show that large-window reasoning can improve small-window decoding.
Oct 27, 2026

Nemotron-Labs-Diffusion-Uni: Elastic Self-Refining Unified Diffusion LLM with Cross-Paradigm On-Policy Distillation

We propose Nemotron-Labs-Diffusion-Uni (NLD-Uni), a unified discrete diffusion large language model (DLLM) for visual understanding and generation tasks. Compared with prior unified diffusion models, NLD-Uni introduces two major innovations. First, we propose a unified cross-paradigm on-policy distillation (OPD) framework that successfully transfers the visual understanding ability of autoregressive VLMs and visual generation ability of continuous flow models, to a unified DLLM. Second, NLD-Uni incorporates a novel progressive self-refining diffusion process that gradually corrupts each token across multiple steps during training, as opposed to adopting the conventional masked diffusion formulation that directly corrupts a clean image token to an uninformative mask token. At inference, this design allows NLD-Uni to make intermediate commit to a token, enabling progressive self-refinement for better visual quality. Through extensive experiments, we show that NLD-Uni achieves state-of-the-art performance among unified DLLMs on diverse tasks including visual question-answering, multi-modal reasoning, object grounding, text-to-image generation, and image editing.
Oct 26, 2026

From Denoising Dynamics to Endpoint Uncertainty in Diffusion Language Models

Predictive uncertainty in stochastic language-model inference can be characterized by repeatedly sampling an input and measuring the distribution of outputs. This requires multiple complete generations. Masked diffusion language models expose another source of information: the evolution of the output during denoising. We ask whether convergence within diffusion trajectories is associated with uncertainty across their final outputs, or endpoints. We study four masked diffusion language models on four reasoning benchmarks, collecting 20 generations per question and 32,000 denoising trajectories in total. Within each trajectory, we measure within-run concentration as the area under the curve describing stable resolution of the eventual output. Across repeated trajectories, we measure endpoint-distribution uncertainty. Stronger within-run concentration is associated with lower endpoint entropy after accounting for model and dataset (β=−0.222β=-0.222, p<10−25p<10^{-25}). This association persists for multi-token outputs and when reweighting positive-entropy cases. The relationship is asymmetric: high within-run concentration almost exclusively accompanies concentrated endpoint distributions, whereas lower concentration occurs under both concentrated and competing endpoint regimes. This relationship strengthens as denoising proceeds. Features from a single denoising run predict whether independent held-out runs will concentrate on a common endpoint or produce competing endpoints. These results connect dynamics within individual denoising runs to variability across repeated inference. Within-trajectory dynamics contain information about whether independent repetitions will concentrate on a common endpoint or produce competing endpoints, providing a route towards estimating endpoint-distribution properties with fewer complete generations.
Oct 26, 2026

Spherical Brownian Diffusion for Language Modeling

Continuous diffusion language models involve two coupled design choices: how to corrupt continuous token representations and how to invert that corruption at generation time. Recent work has explored language generation on the unit sphere through geodesic and von Mises--Fisher (vMF) paths, raising the question of which forward process to use on this geometry and how much that choice matters relative to inference. We introduce Spherical Brownian Diffusion, using Brownian motion as the canonical isotropic diffusion on the sphere. Its heat semigroup provides a principled forward process, an additive dimension-free noise clock, and an exact ancestral bridge, while a clean-token posterior learned with cross-entropy supports multiple samplers. To separate corruption from inference, we derive high-dimensional approximation rates between Brownian and moment-matched vMF corruption and use this relationship for controlled substitutions. On Sudoku, Brownian and moment-matched vMF give nearly identical results across four samplers and five budgets under otherwise identical settings. In contrast, with the corruption process fixed, sampler choice produces a 25.0-point spread at a sampler budget of 180 calls. Using the ancestral bridge, we split this gain into a memory effect from re-noising the state, present on Sudoku and GSM8K, and a token-selection effect on GSM8K and largest at small budgets. Across reasoning tasks, categorical refinement improves performance, whereas unconditional text generation benefits from stochastic updates that preserve diversity.
Oct 25, 2026

Birth, Death and Diff: Insertion-Based Sequence Generation with Corrections

Insertion-based masked diffusion models enable flexible-length, parallel generation by interleaving mask insertion with token unmasking. However, because their generation decisions are irreversible, aggressive parallel decoding can introduce errors that subsequent steps cannot correct, necessitating conservative decoding with more forward passes (NFEs). Moreover, they exhibit higher entropy than standard MDMs because they marginalize over multiple alignments during training and therefore require even more conservative decoding in the absence of a correction mechanism. To address this limitation, we introduce δδMDM, which makes intermediate generations revisable by equipping an insertion-based model with substitution and deletion operations. During training, we expose the model to errors caused by aggressive unmasking and insertion, teaching it to correct its own mistakes. At inference time, we interleave insertion and unmasking with correction steps that can revise previously unmasked tokens and remove unnecessary masks. This mechanism makes aggressive decoding practical, improving both generation speed and quality. δδMDM Pareto-dominates FlexMDM, a vanilla insertion-based MDM, on the speed--quality trade-off. On high-diversity unconditional text generation on OpenWebText, it halves the generative perplexity relative to FlexMDM at the highest inference budget. It also achieves lower generative perplexity at all inference budgets compared to MDM trained from scratch with variable-length padded sequences. On low-diversity conditional text generation on TinyGSM, it matches FlexMDM accuracy with half as many NFEs.
Oct 24, 2026

How To Build A Modern Diffusion Language Model

The frontier of large language model (LLM) performance has advanced rapidly in recent years, fueled by scaling laws for pretraining on large text corpora. However, mainstream LLMs have fundamental limitations: they generate text autoregressively, producing outputs sequentially from left to right one token at a time. This sequential process leaves parallel GPU hardware underutilized during inference--a bottleneck that has grown acute as progress shifts toward scaling post-training and inference-time compute--and commits irreversibly to each token without the ability to revise earlier mistakes. Discrete diffusion models address these limitations by generating tokens in parallel, starting from a noisy or masked sequence and iteratively refining the entire output over a series of denoising steps. This parallel generative process enables speeds of thousands of tokens per second, provides a new axis of inference-time compute scaling, and supports built-in error correction and controllable generation. Over the past two years, discrete diffusion has scaled to billions of parameters across biological sequence modeling and natural language, with open-weights and commercial releases from leading labs. This article provides a survey of key ideas in discrete diffusion and a high-level introduction to how modern diffusion LLMs are built, covering their foundations, core algorithmic and architectural building blocks, applications in biology and language, and implications for future LLM scaling laws.
Oct 23, 2026

LLaCO: Continuous Diffusion for Large Language Models with Competition-Aware Semantic Scheduling

Continuous diffusion models have emerged as a promising non-autoregressive paradigm for text generation over continuous embedding spaces. Along the continuous trajectory, the residual token identity directly shapes the denoising objective. However, the geometric trajectory misaligns with temporal progression, triggering severe temporal compression. In this paper, we show that embedding norm decay causes Euclidean distance to yield spurious overtaking, indicating that token competition must be evaluated directionally. In this space, a candidate overtakes the source token only when the effective noise magnitude surpasses their semantic separation. We formalize these pairwise dynamics as the Competition Threshold Theory (CTT). Crucially, high-dimensional overtaking requires extreme noise levels, causing standard trajectories to concentrate identity decay into a narrow window near the noisy endpoint. To resolve this, we propose Semantic Retention Scheduling (SRS) to align time allocation with the progression of identity decay, and demonstrate that existing token identity metrics are fundamentally statistical aggregations of these pairwise crossings. Building on these theoretical insights, we introduce LLaCO, a generative framework powered by competition-aware semantic scheduling. LLaCO adapts a pretrained backbone to continuous representations via an embedding directional adapter and timestep-conditioned modulation. To operate within the directional space, LLaCO employs a spherical noising process formulated as a geodesic interpolation toward a mask-biased endpoint, eliminating norm decay while preserving accurate token competition. Guided by SRS, LLaCO reallocates the denoising trajectory proportionally to the logarithmic rank decay. Instantiated as an 8B model, LLaCO outperforms previous continuous diffusion baselines and rivals discrete diffusion models of comparable scale.
Oct 23, 2026

Watermarking Diffusion Language Models with Learned Foresight and Balanced Bands

Green-red watermarks favor tokens from a green list that a secret key derives from the preceding tokens. Diffusion language models, however, often choose a token before its preceding tokens, and the token later shapes its neighbors' green lists. Prior work therefore tilts each choice to raise the expected number of green tokens at its own and its neighbors' positions. Building on this objective, we first learn two components that prior work keeps fixed: vocabulary bands that serve as green lists, separate competing candidates and keep the null distribution exact, and a foresight predictor that estimates how each candidate shifts its neighbors' predictions. Second, we count green tokens as the detector does, once per distinct context-token pair, so repetition earns nothing. On WaterBench with LLaDA-8B-Instruct, each learned component raises the z-score at equal distortion, and their gains are nearly additive. With a strong watermark, ours detects at least 96% of responses at a 1% false-positive rate while matching the LLM-judge score of unwatermarked text. Ours also detects more responses on Dream and as many or more on code and question answering without a significant loss in task accuracy.
Oct 23, 2026

DELA: Learning Controllable Edits for Diffusion Language Models

Diffusion language models (DLMs) support bidirectional denoising and iterative refinement, but generation remains largely controlled by token-level masking schedules. This leaves draft-level refinement implicit and limits explicit control over when to insert missing content, delete redundant spans, revisit uncertain regions, or stop editing. We introduce Diffusion Editing Language Adapter (DELA), a plug-in framework that enables controllable draft refinement with a frozen DLM. DELA operates on clean, variable-length drafts using four explicit editing actions: Insert, Delete, Remask, and Stop. The controller determines when and where to edit the draft, while the frozen DLM fills masked regions through denoising, separating high-level edit control from token-level generation. A naive single-phase editor would be computationally expensive because arbitrary interior edits can disrupt prefix stability and limit reliable KV-cache reuse. To address this, DELA uses a two-phase decoding procedure: a fast, append-only drafting phase that supports KV-cache reuse, followed by a global refinement phase that enables targeted non-monotonic edits to the full draft. We train the controller using a staged curriculum that begins with learning local edit priors from synthetic recovery tasks and mined actions that improve draft quality. Subsequent stages refine edit selection through rollout-based outcome preferences and alignment with downstream task feedback. Experiments on mathematical reasoning and code generation show that DELA achieves gains of up to 5.5 percentage points over direct diffusion decoding while keeping the backbone frozen. DELA thus provides a modular approach to explicit draft-level control in diffusion language models.
Oct 22, 2026

What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models

Masked diffusion language models generate tokens in parallel, but their efficiency depends on which predictions are committed. Learned unmasking policies improve the speed--accuracy trade-off, yet differ in supervision, inputs, and capacity, obscuring which choices matter. We view supervised policies as distilling a decoding procedure through a teacher and examine independently what the teacher labels as committable, on which states the policy imitates it, what the policy observes, and how much the number of function evaluations (NFE) must decrease to offset policy overhead. We find that labeling predictions matching the final output is overly permissive at low NFE, and hidden states are weak alone but can complement confidence. About a 10% NFE reduction suffices to offset policy overhead for policies up to 7.5M parameters across tested settings. We therefore train a policy on trajectory-contraction targets, which accept only commitments staying on a reference trajectory, using confidence and hidden states. On LLaDA-8B-Instruct, the resulting 856K-parameter policy achieves the best accuracy--NFE frontier and highest throughput on GSM8K, is competitive at moderate targets on MATH-500 and HumanEval, and largely transfers to Dream-7B-Instruct.
Oct 22, 2026

Harmful Responses Are Jailbreak Templates: Transferring Autoregressive Jailbreaks to Diffusion Language Models

Diffusion large language models (dLLMs) enable flexible masked generation through bidirectional context modeling, but this capability also introduces a distinct jailbreak attack surface. We find that a harmful response to a given query naturally serves as an effective jailbreak template: simply masking a subset of its tokens can induce a dLLM to reconstruct harmful content. Based on this observation, we propose RMT-Jailbreak, a general framework that transfers existing autoregressive (AR) jailbreak attacks to dLLMs through their induced harmful responses. RMT-Jailbreak converts an AR-generated harmful response into a high-mask-ratio template using uniform token retention, and further applies prompt-based safety filtering to mask any remaining visible harmful content. The refined template is then provided to the target dLLM for bidirectional reconstruction. Extensive experiments across multiple AR jailbreak methods and target dLLMs demonstrate the effectiveness and generality of RMT-Jailbreak, revealing strong cross-paradigm transferability of jailbreak attacks.
Oct 22, 2026

Trion-1: Unifying Language, Protein Sequence, and Structure with Diffusion Language Models

We envision a general-purpose protein foundation model as a unified, interactive model of sequence, structure, and function, capable of understanding, generation, and cross-modal translation while retaining the instruction-following and reasoning capabilities of modern LLMs. Toward this vision, we introduce Trion-1, the first native, unified, multimodal diffusion language model (dLLM) for proteins. Trion-1 is built on two key design choices: (1) a unified discrete token space that jointly represents sequence, structure, and natural language; (2) a dLLM backbone that replaces strict left-to-right factorization with bidirectional or semi-bidirectional token interactions, enabling global residue communication and long-range contact modeling that are naturally suited to structure modeling. This formulation unifies six core capabilities: folding, inverse folding, sequence understanding, structure understanding, sequence design, and structure design, while retaining the native language capabilities of general-purpose LLMs. On established benchmarks, Trion-1 achieves broadly competitive performance. In folding, Trion-1 achieves a TM-score above 0.8 with a single sample under strict homology control, matching or exceeding the existing tri-modal model using 200 samples. For sequence understanding, Trion-1 performs on par with top-tier LLMs such as Claude Opus 4.7 (62.37%). These results support well-aligned sequence-structure-language representations and strong cross-modal generative capabilities.