In generative models, steering is often done using knobs—prompts, guidance scales, and property tags. However, adjustments to these knobs will cease to affect the property of interest after a certain point. This limitation is not a flaw of the model but a budget set by the training data before the model is trained.
The movable range of a property splits into two parts: one accessible by the knob and a second, significant part that can only be reached through examples—concrete instances of what you want more of. This second part is usually much larger but not always, and the same budget indicates this in advance.
To reach that second part, a different approach is required: instead of turning a knob, you show the model examples composed of what it has already learned, rather than adding to its training data. A simple audit of the training data measures the budget; we provide a recipe for constructing the example set that accesses all of it.
This approach buys two advantages that a knob cannot:
- Reach: It moves a property across the whole budget, not just the part reachable by a knob.
- Expressiveness: It steers towards targets that can only be specified by example, including those that cannot be articulated in words.
We translate these into a handful of falsifiable claims and verify them in two unrelated domains: image generation and crystal-structure generation, marking where a knob suffices and where only examples will do.
Blogger's Review: This article delves into the steering methods of generative models, proposing the concept of guiding the model through examples. It demonstrates the advantages of this approach in practical applications, offering fresh insights for model optimization that deserve attention!