COFFEE preprint proposes sequence-level guidance for discrete diffusion models

A preprint submitted to arXiv on 28 September 2026 describes COFFEE, a method for steering discrete diffusion models with objectives that apply to a whole sequence rather than to single tokens. The paper, "Grab a Coffee: Future-Aware Guidance for Discrete Diffusion with Compiled Objectives," lists Hua (Edward) Xu, Dongxin Li, Gwen Yidou-Weng, Guy Van den Broeck, Wei Wang and Anji Liu as authors.

Discrete diffusion models generate sequences by resolving several tokens in parallel instead of strictly left to right. The authors say this makes sequence-level guidance hard, because the value of one unresolved token depends on the other tokens it could combine with to form a high-reward sequence. They add that enumerating all such completions makes the computation grow exponentially with the number of unresolved positions.

COFFEE, which the authors call a plug-and-play framework, is meant to avoid that enumeration by separating sequence dependence from the objective. At each diffusion step, a "target-free carrier" takes in the marginal token distributions predicted by the denoiser and builds a joint model over the unresolved tokens. A compiled finite-state model separately records how combinations of those tokens affect the sequence-level preference. Pairing the two models' states, the authors say, lets COFFEE pass global preferences to unresolved positions and sample a clean reconstruction without retraining the diffusion model.

According to the abstract, the same framework handles explicit hard constraints and learned soft objectives. The authors report evaluating it on symbolic, language and biological benchmarks, where they say it achieves strong control results, with quality and diversity trade-offs that depend on the task.

What is confirmed and what is not: the arXiv page confirms the paper's existence, authorship, date and the authors' own description of the method. It gives no benchmark scores, so the size of the reported gains and trade-offs cannot be judged from it. This item draws on a single source, the arXiv abstract page. We found no independent coverage or replication, and an arXiv preprint has not necessarily been peer reviewed. We did not read the full paper.

Why it matters: if the approach holds up, it would let developers apply sequence-level requirements at inference time to a pretrained discrete diffusion model, rather than only checking outputs afterward. The authors frame this as showing the potential of neural-symbolic methods for diffusion guidance. Whether it works as claimed across tasks is for later evaluation to show.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.