COFFEE
A plug-and-play guidance framework for discrete diffusion models that operates through cooperation between an objective-free carrier and a compiled finite-state model. It decouples sequential dependencies from objectives, supports unified control of hard constraints and soft targets, and requires no retraining of pretrained models.
Timeline (last 90 days)
COFFEE is a plug-and-play framework that adds sequence-level goal guidance to discrete diffusion models without retraining the pretrained model.
COFFEE's key innovation lies in decoupling sequence dependencies from the goal itself.
COFFEE introduces two components at each diffusion step: a target-free carrier and a compiled finite-state model.
The target-free carrier is responsible for absorbing the marginal token distributions predicted by the denoiser and constructing a joint model covering all unresolved tokens.
COFFEE transmits global preferences to local unresolved positions and directly samples clean reconstruction results by pairing the states of the two components.
COFFEE can uniformly handle both explicit hard constraints and learned soft goals as control requirements.
COFFEE was evaluated on benchmarks across three domains: symbolic, linguistic, and biological.
COFFEE achieved strong control performance on these tasks and exhibited task-dependent quality-diversity trade-offs.
COFFEE involves sequence-level goals in generation decisions at inference time, whereas the traditional approach only uses goals to score and filter outputs at evaluation time.
COFFEE introduces three new capabilities: joint conditioning, completion-weighted guidance, and optimization-based constraints.
1 more timeline events
All Facts (11)
COFFEE is a plug-and-play framework that adds sequence-level goal guidance to discrete diffusion models without retraining the pretrained model.
50%UnverifiedCOFFEE's key innovation lies in decoupling sequence dependencies from the goal itself.
50%UnverifiedCOFFEE introduces two components at each diffusion step: a target-free carrier and a compiled finite-state model.
50%UnverifiedThe target-free carrier is responsible for absorbing the marginal token distributions predicted by the denoiser and constructing a joint model covering all unresolved tokens.
50%UnverifiedCOFFEE transmits global preferences to local unresolved positions and directly samples clean reconstruction results by pairing the states of the two components.
50%UnverifiedCOFFEE can uniformly handle both explicit hard constraints and learned soft goals as control requirements.
50%UnverifiedCOFFEE was evaluated on benchmarks across three domains: symbolic, linguistic, and biological.
50%UnverifiedCOFFEE achieved strong control performance on these tasks and exhibited task-dependent quality-diversity trade-offs.
50%UnverifiedCOFFEE involves sequence-level goals in generation decisions at inference time, whereas the traditional approach only uses goals to score and filter outputs at evaluation time.
50%UnverifiedCOFFEE introduces three new capabilities: joint conditioning, completion-weighted guidance, and optimization-based constraints.
50%UnverifiedCOFFEE is a paper published on arXiv.
50%