Researchers introduce coffee framework for discrete diffusion model guidance
A new framework called COFFEE addresses a challenge in discrete diffusion models: guiding their generation process with sequence-level objectives. Traditional methods struggle because the value of one unresolved token depends on others, leading to exponential computation growth. COFFEE avoids this by separating sequence dependence from the objective, using a carrier to model unresolved tokens…
Key points
- COFFEE separates sequence dependence from objectives to avoid exponential computation growth in discrete diffusion models
- Uses a carrier model for unresolved tokens and a finite-state model for preference encoding without retraining
- Supports hard constraints, soft objectives, and achieves strong control results across multiple benchmarks
The authors evaluate COFFEE across symbolic, language, and biological benchmarks, reporting strong control results with task-dependent trade-offs between quality and diversity. The framework supports explicit hard constraints and learned soft objectives, enabling joint conditioning, completion-weighted guidance, and optimization-based constraints in pretrained neural generation. This work highlights the potential of neural-symbolic methods in improving diffusion model inference.
Coverage and discussion
2sources- Spectral Feedback for Test-Time Alignment of Protein Diffusion ModelsPrimary source · arXiv cs.AI ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic's Claude begins 'dreaming' to reorganize memory in tests · 1 src
- John Langford quits Microsoft, declares AI Frontiers over · 1 src
- Anthropic releases study on Claude usage in scientific research · 3 src
- MIT and BASF release CheMeleon for low-data molecular prediction · 2 src
- Anthropic tests Claude agents in EVE Online trading and sabotage detection · 1 src
Comments
via GitHub Discussions