Generative Models' Steering Budget: Examples Outperform Knobs.

Generative models are typically controlled using "knobs" like prompts or guidance scales. However, this research reveals a fundamental limitation: past a certain point, these knobs cease to effectively steer the model towards desired properties. The authors introduce the concept of a "steering budget," which is determined by the training data and dictates the total range a property can be moved. This budget is divided into two segments: a smaller portion accessible via traditional knobs, and a significantly larger portion that can only be influenced by providing concrete examples of the desired output. The study demonstrates that this budget can be audited cheaply from the training data, allowing developers to predict in advance when knobs will be insufficient. To access the full steering budget, the paper outlines a recipe for building example sets composed from the model's existing knowledge, rather than requiring additional training data. This "example-based steering" offers greater reach, moving properties across their entire budget, and enhanced expressiveness, enabling steering towards targets that are difficult to articulate verbally. The findings are validated across image and crystal-structure generation domains, providing clear guidance on when to use knobs versus examples.