Reproduce “What's Producible May Not Be Reachable: Measuring the Steerability of Generative Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Keyon Vafa, Sarah Bentley, Jon Kleinberg, Sendhil Mullainathan

Generative models often exhibit poor "steerability," where users struggle to guide them to desired outputs despite high "producibility." This research formalizes a distinction between these two concepts and demonstrates that image-based steering can more than double the user's ability to achieve specific creative goals compared to traditional text prompting.

Key insights

  • Generative models, including state-of-the-art text-to-image and large language models, exhibit consistently low steerability, with human users often unable to achieve desired outcomes.
  • User-driven iterative text prompt refinement provides limited and unreliable improvement, with a notable portion of observed gains attributable to merely trying different random prompts.
  • A tradeoff exists between producibility and steerability, where enhancing one can sometimes come at the expense of the other, and improvements in steerability for newer models are partly due to shifts in their producible sets.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.