cs.CLNov 19, 2026

dLoop: Looped Diffusion Language Models via Recurrent Latent Refinement

Authors: Handa Li, Pengxiang Li, Mingyu Cao, Xilu Wang, Li Shen, Da Li, Lu Yin

Organizations: University of Surrey · The Hong Kong Polytechnic University · Sun Yat-sen University · University of York · 

Abstract

Diffusion language models (DLMs) support parallel generation, yet they often require many denoising steps because only a fraction of token predictions are sufficiently reliable to be committed at each step. Existing acceleration methods primarily optimize confidence learning or commitment policies. However, these approaches operate on the outputs of a single denoising forward pass, in which information from the prompt, previously committed tokens, and uncommitted positions may not yet be fully integrated. Consequently, some predictions remain under-refined when the commitment decision is made, and modifying the commitment policy alone cannot improve their underlying representations. We introduce dLoop, which restructures a pretrained diffusion language model into a Prelude-Loop-Coda architecture and recurrently applies a parameter-shared intermediate block to the same partially masked state. Rather than advancing the denoising process, each additional loop provides further latent computation, allowing information distributed across uncommitted positions to continue interacting and become more fully integrated before commitment. Experiments on GSM8K, MATH, MBPP, and HumanEval show that dLoop achieves the highest or tied-highest accuracy on every benchmark. With tokens per forward pass (TPF) adjusted for loop computation, it outperforms Fast-dLLM in both accuracy and TPF on all four benchmarks and dParallel on three, and is more accurate than d3LLM on all four, while d3LLM attains higher TPF; no compared method surpasses dLoop in both. Mechanistic analyses show that extra loops mainly help on slow, unstable trajectories and that updates at uncommitted positions drive changes in token decisions and later generation. Pre-commitment latent computation thus makes predictions more reliable, offering a route to DLM acceleration distinct from forcing confidence through the training loss.

Figures & tables

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 3, 2026

ALoDLM: Adaptively Looped Diffusion Language Models

Diffusion language models (DLMs) enable fast generation by predicting multiple tokens in parallel, but their practical adoption remains limited by a persistent quality gap relative to comparably sized autoregressive (AR) models. We attribute this gap to a computation-difficulty mismatch: within a partially observed sequence, some unknown tokens are easy to predict, while others require substantially more computation. Existing DLMs nevertheless apply uniform computational depth to all unknown positions at each denoising step. We introduce ALoDLM, which replaces uniform computation with token-adaptive latent recurrence. At each denoising step, ALoDLM iteratively refines latent representations and allocates computation according to token difficulty. Tokens ready to commit are fed back as discrete context, while unresolved tokens retain and further refine their latent states through additional recurrent passes. To learn token prediction and computation allocation jointly, we formulate token-wise computation schedules as latent variables and derive a conditional negative evidence lower bound (NELBO). We train ALoDLM at 1.7B and 8B parameter scales. Across eleven benchmarks, ALoDLM outperforms all evaluated DLMs and the corresponding AR baselines in average benchmark score at both scales. ALoDLM also retains fast parallel decoding, yielding a strong quality-efficiency trade-off among evaluated autoregressive and diffusion models under optimized inference engines.
Oct 21, 2026

Looped Uniform Diffusion Language Models

Uniform diffusion language models (UDLMs) generate text by iteratively denoising sequences of ordinary vocabulary tokens, any of which may be corrupted and subsequently revised. During training, UDLMs must denoise from a context that may itself be unreliable, whereas masked diffusion language models (MDLMs) always condition on a clean context. This challenging training objective makes UDLMs more FLOP-intensive to train than MDLMs, but it yields stronger downstream reasoning performance. Recently, selective looping, which recursively applies a few early Transformer layers within each denoising step, has been shown to improve training efficiency in MDLMs under an iso-FLOP budget. We investigate looping in UDLMs while keeping the training objective and sampler unchanged. We find that looping significantly improves training efficiency: looped UDLMs reach the validation NLL of a non-looped UDLM trained on about 0.5 trillion tokens with up to 2.8x fewer training FLOPs on OpenWebText and 3.3x fewer on FineWeb-Edu. We also find that, at a matched NFE budget, looped UDLMs reach up to 9.8% lower generation perplexity than the non-looped UDLM, with most gains achieved at moderate loop counts. Finally, we study different loop-count distributions and find that sampling smaller loop counts more frequently during training improves inference efficiency: on downstream tasks, this strategy uses 12% fewer loops than uniform loop-count sampling without degrading accuracy.
Jan 5, 2026

CD4LM: Consistency Distillation and aDaptive Decoding for Diffusion Language Models

Diffusion language models (DLMs) accelerate generation by decoding multiple tokens in parallel, but pushing them to fast decoding often sacrifices accuracy. We attribute this to a static-to-dynamic mismatch: training optimizes local denoising under fixed schedules, while fast inference skips steps and encounters irregular mask patterns. Existing accelerators largely address only one side of this problem. Training-free inference methods speed up decoding without changing the model, leaving the mismatch unresolved; training-side distillation methods adapt the model, but often tie the student to a specific decoding trajectory, limiting composability and still suffering accuracy loss. We propose CD4LM, a composable framework for accurate DLM acceleration. Its core training component, Discrete-Space Consistency Distillation (DSCD) trains a trajectory-invariant student on diverse masked states, reducing the mismatch without fixing a decoding policy. We pair DSCD with a lightweight ConfidenceAdaptive Decoder, while keeping the trained student compatible with external inference-side accelerators. On GSM8K, CD4LM preserves accuracy while achieving a 5.18x speedup; stacking CD4LM with per-step cache accelerators further pushes speedup to 9.73x at teacher-level accuracy, suggesting a scalable path for combining training-side robustness with future decoding advances. Code is available at https://github.com/yihao-liang/CDLM.