The Steering Budget: Examples beat Knobs

arXiv cs.AI · 6 hours ago

Researchers found that generative models have a property-adjustment budget determined by training data, where traditional steering methods like prompts and guidance scales can only reach part of this budget, while showing the model concrete examples can access the full range. The study identifies this budget through auditing training data and provides a method for building example sets that reach all of it, with verification conducted across image and crystal-structure generation domains. Using examples instead of knobs enables steering toward targets that cannot be specified in words and achieves full property movement across the entire budget.