COFFEE Enables Future-Aware Guidance for Discrete Diffusion Models

Summary

Discrete diffusion models resolve multiple sequence tokens in parallel, but applying a sequence-level objective is difficult because the value of an unresolved token depends on its possible combinations with other unresolved tokens. Enumerating those completions can make guidance grow exponentially with the number of unresolved positions. The paper introduces COFFEE, a plug-and-play framework that separates sequence dependence from the objective. At each diffusion step, a target-free carrier combines the denoiser’s marginal token distributions into a joint model over unresolved tokens, while a compiled finite-state model records how token combinations affect sequence-level preference. Pairing the two state spaces transfers global preferences to unresolved positions and produces a clean reconstruction without retraining the diffusion model. COFFEE supports both explicit hard constraints and learned soft objectives. Experiments across symbolic, language, and biological benchmarks report strong control, with task-dependent trade-offs between output quality and diversity. The authors present the framework as a way to move joint conditioning, completion-weighted guidance, and optimization-based constraints from evaluation into inference, illustrating the potential of neural-symbolic methods for diffusion guidance.