New AI Method Steers Whole Sentences, Not Just the Next Word

Most AI text generators work like a person writing a sentence one word at a time, always looking back at what they've already written. A different family of AI models, called diffusion models, works more like a crossword puzzle: it starts with blanks and fills them all in over several rounds. That approach powers many image generators, and researchers are now applying it to text, code, and even molecules.

The catch is control. If you want the finished result to satisfy a bigger goal — say, 'this paragraph must not leak private data' or 'this molecule must be non-toxic' — you have to judge each blank by how it affects every other blank. Checking all the combinations explodes fast: 10 blanks with 10 options each is 10 billion possibilities. Most systems simply can't do that in reasonable time.

The new method, called COFFEE, sidesteps this by splitting the job in two. One part tracks the likely word for each blank. The other is a pre-built 'rulebook' — compiled ahead of time — that records how combinations of choices affect the final goal. Pairing them lets the AI push the whole result toward the goal while still filling in blanks, and crucially, without retraining the underlying model. It handles hard rules and softer, learned preferences alike.

The team tested COFFEE on language tasks, logic puzzles, and biological sequences like proteins and DNA. It scored well on control, though quality and variety traded off depending on the task — a familiar tension in AI. This is academic work, not something you can use today. But it points toward a future where you tell a generative AI what you want from the whole result, not just nudge it word by word.