Reproduce “The Geometry of Phase Transitions in Generative Dynamics via Projection Caustics” 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
Ryosuke Sakamoto, Kotaro Sakamoto
Research from Kyoto University and The University of Tokyo proposes that phase-transition-like behaviors in continuous-state generative models stem from projection caustics, where the uniqueness of nearest-point projection onto the data support fails, causing sharp reorientations in the denoising direction. Their Critical Boundary Detector (CBD) identifies these intervention-sensitive windows, enabling efficient, targeted control in models like DDPMs and Stable Diffusion.
Key insights
- Generative "phase transitions" originate from the local geometry of the data support, specifically near projection caustics where multiple data branches compete for nearest-point projection.
- This multi-branch competition manifests as a log-sum-exp structure in the free energy landscape, causing the normalized score direction to rapidly reorient in critical regions.
- The Critical Boundary Detector (CBD) provides a practical and portable diagnostic for identifying these dynamically unstable and intervention-sensitive windows in individual generative trajectories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.