COFFEE framework adds sequence-level guidance to discrete diffusion

Researchers have introduced COFFEE, a framework for steering discrete diffusion models using objectives that apply to whole sequences. The work, titled “Grab a Coffee: Future-Aware Guidance for Discrete Diffusion with Compiled Objectives,” was submitted to arXiv on 28 Sep 2026. It presents an approach for applying preferences during generation rather than assessing them only after a sequence has been produced.

Discrete diffusion generates sequences by resolving several tokens at once, rather than producing them strictly from left to right. That parallel process makes sequence-level guidance challenging: whether an unresolved token is useful can depend on the other tokens that ultimately appear alongside it. The paper says that directly considering every possible completion causes guidance computation to grow exponentially as the number of unresolved positions increases. This is the scaling problem COFFEE is designed to avoid.

COFFEE separates the model of possible token combinations from the objective used to judge them. At each step, a “target-free carrier” takes in the denoiser’s marginal token distributions and uses them to form a joint model over unresolved tokens. Alongside it, a compiled finite-state model tracks how combinations of those tokens relate to the sequence-level preference. The framework pairs the two models’ states to pass global preferences back to unresolved positions and produce a clean reconstruction. The authors say this guidance does not require retraining the diffusion model.

The researchers say COFFEE can handle both explicit hard constraints and learned soft objectives. They evaluated it on symbolic, language, and biological benchmarks, reporting strong control results while noting trade-offs between quality and diversity that depend on the task. The abstract does not give specific benchmark scores or describe a single quality-diversity outcome across all tasks.

For AI builders, the proposal is a way to bring sequence-level requirements into inference for a pretrained discrete diffusion model, rather than relying only on token-level predictions or checking outputs afterward. The paper frames the method as supporting joint conditioning, completion-weighted guidance, and optimization-based constraints. Its results point to neural-symbolic methods as a possible route for controlling diffusion generation, but the reported trade-offs mean builders would need to consider the needs of each task when assessing the approach.