The paper investigates two key design levers for Continuous Diffusion Language Models (CDLMs): self‑conditioning and embedding geometry. It shows that self‑conditioning mitigates a train/inference mismatch and can be implemented efficiently via a two‑level interpolant called DoubleCat, which significantly boosts TinyGSM accuracy. By adding discrete reveals to DoubleCat, the authors create SemiCat, achieving discrete‑diffusion‑level performance and outperforming discrete diffusion on several likelihood‑based benchmarks at 1.7B parameters. The study also finds that cross‑entropy training drives embeddings toward a geometry with decoding error rates comparable to one‑hot representations, and briefly explores additional options such as hybrid corruption processes and asynchronous noise schedules.
arXiv:2605. 00161v2 Announce Type: replace Abstract: Diffusion language models (DLMs) are an attractive alternative to autoregressive models because they promise sublinear-time, parallel generation, yet practical gains remain elusive as high-quality samples still demand hundreds of refinement steps.
The paper introduces a lightweight single‑layer sampler that allows masked diffusion language models to approximate joint sampling of multiple tokens in a single full‑model forward pass. By training the sampler to mimic exact joint sampling from a frozen diffusion model, the authors enable parallel unmasking of tokens while maintaining a close match to the true joint distribution. Experiments on Dream‑7B and Llada‑7B models show that unmasking four tokens per denoising step yields a MAUVE score of 0.87, a substantial improvement over the marginal baseline of 0.31.
The paper introduces LAW (Loopholing All the Way), a continuous generative model that relies entirely on Self‑Conditioning (SC) to propagate expected embeddings across diffusion steps, decoding tokens only once at the end. LAW outperforms both discrete and continuous diffusion models on Sudoku, Maze, and TinyGSM benchmarks, achieving higher accuracy and solving a larger proportion of GSM8K problems. Fine‑tuning a pre‑trained Uniform Diffusion Model with LAW significantly boosts its performance, and on OpenWebText LAW approaches autoregressive models while maintaining a favorable Pareto frontier.
On-Policy Residual Learning for Diffusion Language Models (ORL) introduces a framework that learns residual context directly from model-generated trajectories, using an entropy-contraction objective to sharpen predictions. The approach co-designs algorithmic and system components—such as asynchronous residual capture, block-diffusion attention, and packed compute-balanced training—to enable practical scaling. Experiments on a 4B DLM converted from Qwen3-4B show up to 6.2× faster tokens per forward and 6.9× higher end-to-end throughput, while also improving speed-quality trade-offs over autoregressive models and transferring effectively to other DLMs like LLaDA-2.1-mini and DiffusionGemma.
arXiv:2609. 04010v1 Announce Type: new Abstract: Large Language Models (LLMs) owe much of their success to next-token prediction (NTP), but their autoregressive (AR) structure requires slow, sequential token generation.
The paper introduces the Language Diffusion Transformer (LangDiT), a new transformer architecture tailored for continuous diffusion language models (CDLMs). By incorporating feature and query-key normalization, wider noise conditioning, likelihood-shortcut gradient routing, calibration, and confidence-gated training self-conditioning, LangDiT achieves lower generation perplexity than existing models across various CDLM formulations. Experiments on the OpenWebText dataset show state-of-the-art perplexity scores, demonstrating the architecture’s effectiveness while preserving the underlying model representations and training objectives.
PlaidQ is a 0.7B continuous diffusion language model designed for code generation. By distilling its iterative refinement trajectory into only a few denoising steps—or even a single step—PlaidQ achieves competitive performance with discrete diffusion models while dramatically reducing inference time. The study demonstrates that continuous diffusion can be effectively compressed, enabling efficient and accurate code generation with minimal computational overhead.
arXiv:2607. 14106v1 Announce Type: cross Abstract: In this paper we introduce token time continuous diffusion (TTCD), a new diffusion language model which (a) operates in continuous space, deterministically mapping Gaussian noise to a final token canvas with no further sampling, and crucially (b) incorporates a new notion of per-token times, with some tokens proceeding from noise to token at a faster rate than others.
arXiv:2607. 24507v1 Announce Type: cross Abstract: Existing methods mainly adapt pretrained autoregressive (AR) language models to masked diffusion, whereas we directly adapt them to uniform-noise diffusion, where every token remains editable during sampling.
The paper introduces Sigma, a large-scale continuous diffusion language model (3B/8B parameters) that uses steerable, low-dimensional ODE/SDE latent trajectories to address the non-smoothness of discrete diffusion models. Sigma is trained blockwise via likelihood optimization, jointly denoises Gaussian-corrupted token embeddings, and learns an optimal embedding geometry, leveraging pre-trained autoregressive weights for faster training. During inference, classifier-free guidance and score temperature are identified as essential for high-fidelity reasoning and coding, and Sigma matches or exceeds discrete models on benchmarks such as GSM8K, Minerva, HumanEval, MBPP, MATH-500, and AIME, while also revealing unique structural benefits like embedding-space steering and graceful degradation for low NFEs.
arXiv:2505.16990v3 Announce Type: replace Abstract: In this work, we propose Dimple, the first Discrete Diffusion Multimodal Large Language Model (DMLLM). We observe that training with a purely discr...
arXiv:2609.37974v1 Announce Type: cross Abstract: Masked diffusion models (MDMs) generate text by unmasking several tokens per step, but they are trained and sampled under different conditions. The m...