Predict, Don't Iterate: Efficient Adaptive-Length Infilling for Diffusion Language Models

Diffusion language models (DLMs) have emerged as a promising alternative to the auto-regressive paradigm. With bidirectional attention and any-order generation, DLMs naturally fit infilling tasks, which require generating a middle span conditioned on both the prefix and the suffix. However, infilling is sensitive to the length of the span, while DLMs require the length to be fixed before generation. Although prior studies extend DLMs to dynamic lengths, they still suffer from two limitations. (i) Sensitivity to initial length. These methods require a preset length to initialize the search and are highly sensitive to this initial length, often yielding suboptimal results. (ii) Inference inefficiency. They either insert length-changing operations during generation or repeatedly search for an appropriate length using multi-step denoising confidence, both of which introduce substantial extra forward passes and computational cost. Therefore, we propose PILL (Probing-based InfiLling with preset-Length-free decoding), an efficient infilling method for DLMs that requires no preset initial length and adds far fewer extra forward passes than baselines, substantially reducing inference time. Experiments show that, across five DLMs spanning different families, architectures, and training recipes on eight infilling benchmarks, PILL improves over the strongest baseline by +4.8 average pass rate on code and +6.0 BLEU-2 on text, while running 1.82x faster than that baseline. The code is available at https://github.com/Hsu1023/PILL.

hLLM: Single Pass Decoding for Generative Reranking

Large language models (LLMs) achieve state-of-the-art generative ranking quality, but the ranking they produce must be decoded, and autoregressive decoding spends one sequential forward pass per emitted token. We observe that the only tokens a ranker must emit are the $N$ ordinal values naming the items in ranked order, and that this narrow, permutation-structured output format admits decoding strategies which are much more efficient than left-to-right generation. We introduce hLLM (Hungarian LLM), a format-specialized decoding strategy that decodes all $N$ ordinals in $O(1)$ forward passes. hLLM reads an $N \times K$ item-position score matrix off the LLM's prefill hidden states with a lightweight self-attention head, then decodes the ordinals as the optimal bipartite assignment of that matrix via the Hungarian algorithm, yielding a valid permutation by construction rather than by repair. Through a systematic study of training signals and backbone adaptation, we show that LoRA-based fine-tuning combined with teacher ranking distillation reaches 28 ms end-to-end inference, a speed-up of $64\times$ while maintaining ranking quality on par with the teacher. We provide a complete ablation decomposing the contributions of architecture, training signal, and backbone adaptation. Our framework connects generative ranking to combinatorial optimization, opening a path toward other $O(1)$-decode mechanisms for real-time ranking.

A Model with No Head and Many Thoughts

Large language models decode by projecting hidden states through a large vocabulary head at every step. This operation is computationally costly and forces all reasoning to be expressed in discrete tokens. We introduce Soft Latent Thinking, a method that replaces the LM head during reasoning with a lightweight projector, enabling autoregressive rollout in embedding space where reasoning steps remain continuous rather than tokenized. Experiments on DeepSeek-Qwen-1.5B and LLaMA-3.2-3B show that Soft Latent Thinking consistently improves pass@k across all k while reducing per-step compute during chain-of-thought. Our method achieves the highest pass@32 among all soft-thinking approaches, demonstrating that effective reasoning can be carried out in continuous space without discrete token generation.

Locate Anything in Videos: Rethinking Efficient Generative Spatio-Temporal Video Grounding

Spatio-temporal video grounding (STVG) requires models to identify when a referred event occurs and localize the target entity throughout that interval. Existing multimodal large language models typically serialize dense localization trajectories autoregressively, causing decoding latency to grow with tube length and allowing localization errors to propagate across time. We introduce Parallel Tube Decoding (PTD), a generative formulation that decomposes grounding into a temporal block followed by time-conditioned spatial blocks decoded simultaneously. This removes both token-level and trajectory-level dependencies, reducing the sequential decoding depth to a fixed $1 + 1$ rounds, independent of tube length. To enable parallel spatial generation, we introduce Decoupled Block Attention, which preserves access to shared video-query context while eliminating cross-box dependencies, together with localization-aware policy optimization for temporal boundaries and spatial geometry. On VidSTG, PTD reduces Tube Completion Latency by 79x and increases spatial decoding throughput by 92x over standard autoregressive decoding, while also improving grounding accuracy. With a compact 4B backbone, our model performs favorably well on VidSTG and HC-STVG, and generalizes zero-shot to temporal grounding, grounded VideoQA, and referring video object tracking. Our results show parallel tube generation is an efficient and effective alternative to autoregressive localization in videos.

Survival-Guided Length Control for Efficient Diffusion Language Models

Diffusion language models (DLMs) generate text by iteratively denoising masked sequences, but standard decoding either fixes the sequence length or relies on ad hoc stopping rules, often leading to unnecessary denoising steps. We recast length selection as a discrete-time survival problem over the end-of-sequence token and propose a plug-in, training-free length predictor that can be added to any existing DLM. Across reasoning and code-generation benchmarks, survival-guided length decoding speeds up inference by up to 7 times while preserving task accuracy. We further find that predicted lengths vary widely even within the same dataset, making model performance sensitive to the chosen length.

DiffusionGemma Technical Report

We introduce DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at exceptionally high speed. Rather than decoding one token at a time, DiffusionGemma iteratively refines blocks of 256 tokens in parallel, avoiding the sequential decoding bottleneck of conventional autoregressive (AR) large language models. Instead of training from scratch, we obtain DiffusionGemma by fine-tuning the mixture-of-experts Gemma 4 model with 3.8B activated and 25.2B total parameters. Our compute-efficient two-stage training pipeline uses fewer than 10% of the starting AR model's total training token budget. The first stage uses supervised fine-tuning to teach bidirectional denoising, while the second stage combines reinforcement learning with sampler distillation to jointly improve generation quality and inference efficiency. DiffusionGemma establishes a new Pareto frontier for the trade-off between generation speed and model capability. Averaged across our full evaluation suite, it generates around 20 tokens per forward pass and achieves roughly 1,500 output tokens per second on a single NVIDIA H100 GPU, which is substantially faster than AR models even with state-of-the-art speculative decoding. DiffusionGemma also retains the starting model's support for thinking mode, multimodal inputs, and long contexts. Despite diffusion fine-tuning, it remains capable of AR generation with only minor performance degradation, suggesting a path toward hybrid diffusion-AR decoding.

DIRECT: Direct Decoding for Efficient and Aligned Sequence Labeling with Large Language Models

Sequence labeling is a fine-grained information extraction task, yet existing large language model-based approaches suffer from insufficient domain alignment and low inference efficiency. To address these issues, we propose DIRECT, a framework that addresses these issues through training-time optimization and inference-time rectification. Specifically, DIRECT performs Direct Preference Optimization (DPO) after supervised fine-tuning to strengthen task alignment with human preferences, and introduces a controlled decoding process that enforces fixed output formats and restricts predictions to candidate sets. To further improve efficiency, a template-filling mechanism requires the model to generate only label tokens while reusing prefixed content through the KV Cache, thus reducing redundant computation. Experimental results on eight datasets demonstrate that DIRECT achieves significant improvements in both performance and efficiency compared to existing methods.

$x$-Prediction Flow: Efficient Continuous Decoding for Masked Diffusion Language Models

Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens, but their standard decoder reduces each step to a binary action: a position is either committed to a single token or left fully masked, discarding rich predictive information rather than carrying it forward, and forcing premature, irrevocable commitments that lead to poor performance under a limited decoding budget. In this paper, we reinterpret mask prediction as a clean-state prediction ($x$-prediction) and show that it can be used to induce a continuous flow in the input embedding space. Building on this view, we propose a continuous decoding framework for MDLMs where tokens can accumulate partial progress at each diffusion step and remain revisable. To match the uneven contextual constraints across positions in language, we replace the globally synchronous schedule in image diffusion with a confidence-based asynchronous update in which the diffusion progress is token-wise accumulated. Additionally, we introduce a lightweight policy network and formulate its training as a reinforcement learning problem. Applied to pretrained LLaDA, our decoder retains 83--97% of full-budget accuracy using under 15% of the diffusion steps, largely outperforming discrete mask-prediction decoding at matched budgets.

GLIDE: Guided Layerwise Hybrid Attention for Efficient LLM Inference

As Large Language Models scale to increasingly long contexts, the memory I/O and computational overhead of the Key-Value (KV) cache during decoding emerges as the primary throughput bottleneck. To address this, we propose GLIDE, a Guided Layerwise Hybrid Attention that strategically integrates sliding-window softmax attention with linear recurrent aggregation. GLIDE is motivated by layer-wise heterogeneity: early layers exhibit high sensitivity to softmax removal, while deeper layers demonstrate redundancy and tolerate aggressive replacement by linear alternatives. Leveraging this insight, GLIDE introduces a layer-wise adaptive mechanism wherein each layer balances an efficient linear recurrence with a variable-sized softmax window. Unlike uniform hybrid approaches, GLIDE non-uniformly compresses the softmax footprint across the model, reducing aggregate KV cache I/O while preserving expressive power where most vital. Empirical evaluations demonstrate the GLIDE achieves superior performance-efficiency tradeoffs, reducing end-to-end latency for long-context generation without compromising quality.

DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models

Block diffusion language models accelerate decoding through parallel block-wise denoising, yet whether they can be reliably scaled for long chain-of-thought (CoT) reasoning remains unresolved. To this end, we develop DreamReasoner-8B, an open-source block diffusion reasoning model, and conduct a systematic study of how training and inference block sizes affect long-CoT reasoning. Our analysis reveals a stark performance disparity: training with large block sizes yields remarkably poor reasoning, whereas small block sizes preserve effective reasoning. To bridge this granularity gap, we propose block-size curriculum learning, which gradually transitions training from fine-grained to coarse-grained block sizes, thereby overcoming this limitation and enabling strong reasoning performance that generalizes across diverse inference block sizes. On mathematical and code reasoning benchmarks, DreamReasoner-8B achieves results competitive with leading open autoregressive models such as Qwen3-8B. This work establishes a practical foundation for efficient, reasoning-capable diffusion language models. We release our model at https://github.com/DreamLM/DreamReasoner.

VoidPadding: Let [VOID] Handle Padding in Masked Diffusion Language Models so that [EOS] Can Focus on Semantic Termination

MDLMs generate text by denoising a preallocated masked response canvas, making response-length modeling central to instruction tuning. Existing MDLMs often inherit the autoregressive convention of using repeated \texttt{[EOS]} tokens for padding during instruction tuning, giving \texttt{[EOS]} a dual role as both a semantic terminator and a padding token. We show that this dual role is a root cause of \texttt{[EOS]} overflow under large-block decoding. To decouple these roles, we propose VoidPadding, which introduces \texttt{[VOID]} for padding and reserves \texttt{[EOS]} for termination. During inference, the learned \texttt{[EOS]} signal enables early stopping, while the learned \texttt{[VOID]} signal guides adaptive response canvas expansion. On Dream-7B-Instruct, VoidPadding improves the block-size-averaged four-task mean across mathematical reasoning and code generation benchmarks by \(+17.84\) points over the original model and \(+6.95\) points over RainbowPadding, while reducing decoding NFE by 55.7\% on average. Code is available at https://github.com/Haru-LCY/VoidPadding.

SegDINO: Introducing Multi-Scale Structure into DINO for Efficient Medical Image Segmentation

Self-supervised DINO models provide strong transferable visual representations, yet applying them directly to image segmentation remains challenging. Existing approaches commonly rely on heavy decoders with complex upsampling, introducing substantial parameter and computational overhead. We observe that introducing scale into DINO features is far more critical than increasing decoder capacity. In this work, we present SegDINO, an efficient segmentation framework that integrates a DINOv3 backbone with lightweight scale modeling. SegDINO introduces Token Pyramid Adaptation (TPA) to reorganize intermediate DINO features into a pseudo multi-scale hierarchy, and Scale-Aware Decoding (SAD) for efficient intra-scale refinement and top-down multi-scale propagation. We further curate PanCT, a new CT dataset containing 284 patients with expert-annotated pancreatic tumors, to assess SegDINO's ability to handle difficult small-lesion cases. Extensive experiments on PanCT and three public benchmarks demonstrate that SegDINO achieves state-of-the-art results with high efficiency. The code is available at https://github.com/script-Yang/segdino_v2.

Ghost Attractor Networks: Basin-Structured Dynamical Decoders for Closed-Loop Sequential Generation

Sequential output generation with large-scale Transformer and diffusion decoders pays a memory cost that grows with sequence length, plus iterative per-step computation. Replacing them with small feed-forward decoders restores efficiency but produces unstructured latent representations that limit closed-loop control: phase-conditioned action generation and cross-step latent carry-over both require a latent geometry with stable basins. This article proposes Ghost Attractor Networks, a theoretically derived dynamical decoder whose latent evolves under a learned potential with drift and produces a basin-attractor structure by construction. Three desiderata (multi-modality, decoder-level single-pass switching, and constant memory) motivate the potential-drift form, and mode transitions arise as saddle-node bifurcations with ghost-attractor escape. A hierarchical phase-space decomposition separates first-order basin convergence from second-order proprioceptive refinement. Empirically, a Ghost trained end-to-end with a behavioral-cloning and contrastive objective exhibits the predicted gradient-flow contraction in its potential, with the gradient norm decaying by 67 percent across five integration steps on 1430 held-out samples. Ghost is evaluated as a robotic action decoder. A 2.3-million-parameter Ghost matches the offline accuracy of a 1.07-billion-parameter Diffusion Transformer at 462 times fewer parameters and 32 times lower latency, and beats five alternative 2M-parameter decoders (MLP, Neural ODE, CVAE, Transformer, 1-step Diffusion) on offline mean squared error by 5.9 to 29 percent. On the LIBERO-10 closed-loop benchmark, phase conditioning on Ghost's basin-structured latent yields a 13.5 percentage-point success-rate gain over a feed-forward MLP baseline, and persistent-latent ensembling reaches a 95.7 percent final success rate.

QK-Normed MLA: QK normalization without full key caching

Query-key (QK) normalization stabilizes attention by controlling the scale of queries and keys before the dot product, but is not immediately compatible with Multi-head Latent Attention (MLA). MLA achieves efficient decoding by caching low-dimensional latent states instead of full keys, whereas post-projection QK RMSNorm appears to require the fully projected key for every cached token. We show this apparent incompatibility is an implementation artifact, not an architectural constraint. RMSNorm decomposes into a static affine weight and a dynamic scalar RMS statistic. The static key-side weight can be absorbed into the MLA query-side projection; the dynamic key statistic reduces to one inverse-RMS scalar per token and KV group. The resulting formulation is exactly equivalent to explicit post-projection QK RMSNorm in exact arithmetic and preserves MLA's latent decode path. In our 400M runs trained for up to 100B tokens, QK-Normed MLA achieves lower training loss and better downstream accuracy than QK clipping, while H800 decode benchmarks show less than 2% latency overhead up to 256k context. These results make QK normalization a practical stabilization option for MLA models without requiring full-key caching.

K-Forcing: Joint Next-K-Token Decoding via Push-Forward Language Modeling

Autoregressive (AR) language modeling is the dominant paradigm for text generation, yet its sequential token-by-token decoding makes inference memory-bound and inefficient. Existing acceleration approaches, such as speculative decoding and diffusion language models, can yield speedups under certain conditions but do not directly address high-load batch serving--the scenario most critical for industrial-scale deployment. We introduce K-Forcing, a push-forward language modeling paradigm for joint next-k-token decoding. K-Forcing distills an existing AR model into a conditional push-forward mapping--one that transforms independent uniform noise variables into a joint sample of multiple future tokens in a single forward pass. This design preserves fixed-length outputs, reuses the AR teacher backbone, and remains compatible with standard AR serving infrastructure. We train this mapping via progressive self-forcing distillation, which gradually expands the prediction window while enabling the student to closely match the sequence distribution of the AR teacher. We evaluate K-Forcing on LM1B and OpenWebText using a standard causal Transformer backbone. When aggressively configured to generate k = 4 tokens per forward pass, K-Forcing delivers approximately 2.4-3.5x speedup across different batch sizes, while incurring modest quality degradation relative to its AR teacher. As inference increasingly dominates the lifetime compute cost of modern LLMs, K-Forcing offers a promising route toward accelerating AR generation under real-world high-load deployment.

SemBlock: Semantic Boundary Dynamic Blocks for Diffusion LLMs

Diffusion language models (DLMs) generate text through iterative denoising, and blockwise decoding improves their practicality by committing tokens in local blocks. However, existing blockwise methods typically rely on fixed block sizes or delimiter-based runtime signals, which do not necessarily align with semantic boundaries. In this paper, we propose SemBlock, a semantic-boundary-driven dynamic block decoding framework for diffusion LLMs. SemBlock formulates dynamic block construction as semantic boundary prediction and trains lightweight predictors on frozen LLaDA hidden states. To provide supervision, we construct SemBound, a semantic-boundary dataset that derives boundary labels from discourse units, reasoning steps, and implementation spans across natural language, math, and code tasks. During inference, SemBlock uses predicted boundary probabilities to select the ending position of each dynamic block. Experiments on GSM8K, IFEval, MATH, and HumanEval show that SemBlock consistently improves over fixed-block decoding and AdaBlock. Our code is publicly available: https://github.com/TH-AI-Lab-PKU/SemBlock.

ChannelTok: Efficient Flexible-Length Vision Tokenization

Leading flexible vision tokenizers achieve SOTA quality at an extreme cost, relying on parameter-heavy backbones and slow, multi-step generative decoders. We depart from this complex, spatial-token paradigm and introduce a simple, lightweight, and fast channel-wise flexible-length tokenizer. Our method treats each latent channel as a visual token, enabling a parameter-efficient CNN-Transformer hybrid backbone. Furthermore, employing a stochastic tail-dropping paradigm during training naturally forces channels to organize by semantic importance. This allows for flexible compression at inference by simply retaining the first $k$ channels, and naturally enables variable-length autoregressive image generation. We validate our approach through extensive experiments on ImageNet, demonstrating consistent quality across diverse token budgets. The results establish a new quality-efficiency frontier: our model achieves state-of-the-art perceptual quality (rFID 2.92) while being $8.6\times$ faster in decoding and $2.1\times$ smaller (159M params) than the next-best alternative. Our work establishes channel-wise tokenization as a powerful and practical paradigm for efficient visual representation. Project page: https://channeltok.github.io

Supportive Token Revealing for Fast Diffusion Language Model Decoding

Discrete diffusion language models can generate text efficiently by updating multiple masked positions in parallel, but this parallelism introduces a quality-latency trade-off. Aggressive decoding may commit mutually dependent tokens too early, while conservative decoding requires many denoising steps. Existing methods address this tension by deciding which tokens are safe to reveal using confidence or dependency criteria. However, avoiding unsafe commits does not necessarily make the remaining masked sequence easy to decode, since uncertain tokens may depend on masked tokens, creating a bottleneck for denoising steps. We propose AXON, a training-free module that can be added on top of existing parallel decoding strategies for diffusion language models. Rather than replacing the base decoder, AXON monitors the remaining uncertain masked tokens and intervenes only when their current state suggests that additional context is needed. It then shifts the criterion from which tokens are safest to reveal to which confident reveals would best support later denoising. AXON selects anchors, confident masked tokens that uncertain positions attend to, using attention, uncertainty, and confidence signals. Experiments on reasoning and code-generation benchmarks across multiple diffusion language models show that AXON improves the quality-latency trade-off of existing parallel decoders, often reducing the number of function evaluations while maintaining or improving accuracy.