Tag
BYOD presents a LoRA-only recipe for efficiently converting autoregressive language models into diffusion language models, achieving conversion of Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B in about 3-4 hours on a single Colab GPU while keeping all base parameters frozen.
The paper measures dynamic evaluation (taking gradient steps on text being read) across softmax-attention, hybrid linear-attention, and masked-diffusion language models, finding 5–20% next-token loss reductions but showing the adaptation captures style/vocabulary rather than factual retrieval, while low-rank distillation of a model's own conclusions reaches in-context performance without degrading GSM8K.
TraceQuery 是一个轻量级稀疏解码框架，通过全去噪轨迹监督预测需要重新计算的少量位置，为扩散大语言模型（dLLM）推理提供最优的质量-计算权衡。
该论文提出通过学习词汇带(bands)和预视预测器(foresight predictor)来改进扩散语言模型的绿-红水印方法，在WaterBench上使用LLaDA-8B-Instruct实现96%以上的检测率，同时保持与无水印文本相当的生成质量。
DTCA introduces a token-level credit assignment framework for reinforcement learning in diffusion language models, measuring token contributions via confidence improvements and entropy calibration. Experiments on LLaDA-8B-Instruct show gains of up to 9.0 points on MATH500 and 5.3 points on GSM8K for mathematical reasoning.
DiffResearch 将自回归骨干替换为扩散语言模型（dLLM）用于多智能体深度研究，在 Deep Research Bench 上实现约 6 倍端到端加速并提升报告质量，还提出 dFast 提前从去噪状态提取子查询，以 Apache-2.0 开源。
该论文研究扩散语言模型智能体之间能否通过压缩的连续隐藏向量进行因果性通信，发现去噪状态的传递能显著影响接收方推理，且单次晚时传递的效果优于多次重复交换。
The paper introduces SPOT, a 0.62M-parameter drafter that predicts token commit order for diffusion language models, enabling lossless tree-based speculative decoding with coupled-noise match verification that cuts target passes by up to 6.8x and boosts throughput up to 3.2x across six DLMs.
该论文提出ORL（On-policy Residual Learning）框架，通过熵收缩目标与异步残差捕获等系统级协同设计，使扩散语言模型（DLM）在相近质量下将吞吐量提升至约6.9×，在单请求解码速度上超越自回归模型的推测解码方案。
提出训练无关的解码算法 One Block Ahead (OBA)，在块扩散语言模型中并发去噪当前块与下一块并传递采样器状态，在 DiffusionGemma 上将每次前向传递的 token 数提升 42.3%、单请求吞吐提升 16.3%。
A paper proposing a likelihood-free post-training pipeline for diffusion language models combining outcome-routed distillation with ratio-free group-centred REINFORCE and a KL anchor, showing large tool-calling gains over a teacher that succeeds on only 18.6% of prompts.
Yohei Nakajima 转发介绍 Next-Concept Modeling (NCM)：一种快速、并行、由粗到细的文本生成新范式，在摘要等下游任务上相比同规模扩散语言模型最快可达 31 倍加速。
The paper proposes Predictive-State Flows (PSF), a continuous diffusion language model design where a rectified-flow prior generates block-level predictive states and a block-causal decoder generates tokens in parallel, significantly reducing diffusion states and sequential depth while improving efficiency and generative quality over models like ELF-B.
This paper proposes Seq-Drifting, a framework that learns a map from Gaussian noise to text in a single forward pass by applying drifting to sequences of continuous token representations, achieving a generative perplexity of 36.01 on LM1B with large speedups over few-step and multi-step diffusion language models.
Next-Concept Modeling (NCM) 提出一种 coarse-to-fine 两阶段框架，利用多尺度残差 VQ-VAE 将文本编码为层级离散码并逐层并行生成，在 LM1B、OpenWebText 与 XSum 上以 10-31 倍加速达到或超越最强扩散语言模型基线，并在 Sudoku 推理任务上显著优于自回归与扩散基线。
论文提出 CoDR（Confidence-Drift Remasking），一种免训练、与采样器无关的后处理精炼方法，通过检测已提交 token 的置信度漂移并对其进行重遮蔽与重新生成，从而在掩码扩散语言模型的推理与编码任务上以较少的前向传播次数提升准确率。
SpecFold 论文提出一种算法-系统协同设计，通过识别并折叠多分支投机验证中的计算冗余（token级残差门控与折叠注意力/FFN复用），在扩散大语言模型上实现最高1.64x吞吐提升（相比Spiffy），且保持任务性能相当。
NVIDIA introduces Sigma, the first large-scale (3B/8B) continuous diffusion language model, which denoises Gaussian-corrupted token embeddings over steerable low-dimensional ODE/SDE trajectories, achieving performance parity with discrete diffusion and autoregressive models on math and coding benchmarks while enabling embedding-space steering for quality-diversity trade-offs and efficient few-step distillation.
SoloQ提出一种免校准的量化框架，将权重和激活映射到归一化旋转基底，实现数据无关的4比特量化，在扩散语言模型上保持精度并最多降低2.61倍峰值内存、加速2.24倍推理。
This paper introduces a post-training method for diffusion language models that minimizes Maximum Mean Discrepancy between generated and reference distributions in the feature space of a frozen pretrained DLM, achieving lower generative perplexity and better accuracy-compute trade-offs on OpenWebText and GSM8K, with improved decoding parallelism on 16B LLaDA-2.0 models.