dLoop restructures diffusion language models with recurrent latent refinement
The arXiv paper proposes looping a shared block over partially masked states so diffusion language models commit tokens with better-integrated representations, reporting accuracy and speed gains over named baselines.
In plain words
Diffusion language models generate text in parallel but often need many cleanup steps because they only commit tokens they consider reliable. The dLoop paper restructures a pretrained model so a shared intermediate block runs repeatedly on the same partially finished draft, giving the model extra internal computation before it commits tokens. The authors report higher accuracy and, adjusted for the extra computation, better throughput than several named acceleration methods.
What was announced
Diffusion language models can generate tokens in parallel, but the paper says they often need many denoising steps because only a fraction of predictions are reliable enough to commit at each step. Prior acceleration work focused on confidence learning or commitment policies, which the authors argue operates on outputs of a single forward pass where prompt, committed, and uncommitted information is not fully integrated, leaving some predictions under-refined.
dLoop instead restructures a pretrained diffusion model into a Prelude-Loop-Coda architecture, recurrently applying a parameter-shared block to the same partially masked state. Each loop adds latent computation rather than advancing denoising, letting uncommitted positions keep interacting before commitment. On GSM8K, MATH, MBPP, and HumanEval, the authors report the highest or tied-highest accuracy on every benchmark; with tokens per forward pass adjusted for loop computation, it outperforms Fast-dLLM on all four benchmarks and dParallel on three, and is more accurate than d3LLM on all four, though d3LLM attains higher TPF. Mechanistic analyses suggest extra loops mainly help slow, unstable trajectories, with updates at uncommitted positions driving later token decisions. These are the authors' reported results, not independently confirmed.
What it does not show
- Evidence is limited to the paper's abstract; full experimental details, model sizes, and setup are not available in the supplied passages.
- Reported comparisons and benchmark results are the authors' own claims and are not independently confirmed.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.