New research proposes 'steering budget' limits in generative models

A new paper introduces the concept of a "steering budget" in generative models, proposing that the range of a model's steerability is limited by its training data. This budget dictates how much a property can be influenced by traditional "knobs" like prompts or guidance scales. The research suggests that to reach the full potential of a model's capabilities, especially for properties that cannot be easily verbalized, concrete examples are more effective than knobs. The authors provide a method to audit this training data budget and construct example sets to maximize steerability, demonstrating their findings in image and crystal-structure generation. AI

IMPACT This research could lead to more controllable and expressive generative models by providing a framework for understanding and maximizing their steerability.

RANK_REASON The cluster contains a research paper published on arXiv detailing a new concept in generative models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• Raj Kumar Rajendran
• ScienceCast
• The Steering Budget: Examples beat Knobs

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →