RESEARCH · RESEARCH · #1253
COFFEE: future-aware guidance for discrete diffusion (arXiv:2609.35924v1)
The paper introduces COFFEE, a plug-and-play inference-time framework that guides discrete diffusion models with sequence-level objectives without enumerating completions. It combines a target-free carrier that absorbs marginal token distributions from the denoiser with a compiled finite-state model that records how token combinations affect the objective, supporting hard constraints and learned soft objectives and reporting strong control results across symbolic, language, and biological benchmarks with task-dependent quality/diversity trade-offs.
KEY POINTS
- The paper introduces COFFEE, a plug-and-play inference-time framework that guides discrete diffusion models with sequence-level objectives without enumerating completions.
- It combines a target-free carrier that absorbs marginal token distributions from the denoiser with a compiled finite-state model that records how token combinations affect the objective, supporting hard constraints and learned soft objectives and reporting strong control results across symbolic, language, and biological benchmarks with task-dependent quality/diversity trade-offs.
- It matters because COFFEE enables efficient sequence-level guidance and constraints for pretrained discrete diffusion generators at inference time without retraining, broadening practical control over generated sequences.
WHY IT MATTERS
It matters because COFFEE enables efficient sequence-level guidance and constraints for pretrained discrete diffusion generators at inference time without retraining, broadening practical control over generated sequences.
SOURCES & TIMELINE
1arXiv:2609.35924v1 Announce Type: new Abstract: Discrete diffusion models generate sequences by iteratively resolving multiple tokens in parallel, offering a flexible alternative to left-to-right generation. However, guiding this process with a sequence-level objective is difficult because the value of one unresolved token depends on the other tokens with which it can form a high-reward sequence. Enumerating all such…
arXiv:2609.38448v1 Announce Type: new Abstract: Open-ended LLM homogeneity can create false plurality when several systems appear to offer independent perspectives while returning the same familiar default. Single-pass answers obscure the distinction between agreement produced by a tightly constrained answer space, prompt-vocabulary echo, and broader answer spaces with stable alternatives beneath the surface. We intr…