Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces Finite-Time Spectral Sensitivity (FTSS), a novel metric that analyzes the geometric properties of flow matching models by tracking the singular values of the state-transition matrix. It reveals that while models that generalize maintain stable effective dimensions, those that overfit exhibit a spectral collapse under data scarcity. This internal geometric audit allows for the detection of generative memorization without relying on external data comparisons, providing a new diagnostic tool for understanding model behavior.
Overfitting in generative models leads to a spectral collapse, exposing a hidden geometry of memorization that can be diagnosed without external data.
Continuous-time generative frameworks construct probability paths between base and target domains by optimizing time-dependent velocity fields. While theoretical targets favor straight trajectories, empirical networks develop complex path deformations. This paper presents the Finite-Time Spectral Sensitivity (FTSS) g(t), a gradient-free, forward-pass metric that exposes flow geometry by tracking the root-mean-square singular value of the state-transition matrix. Serving as a continuous proxy for stable rank, g(t) reveals a distinct geometric pathology under data scarcity: while generalizing models maintain stable effective dimensions, overfitting causes a spectral collapse. We leverage this structural phenomenon to develop an internal geometric audit based on g(t). Our framework detects generative memorization using purely internal trajectory dynamics, removing the need for external membership queries or baseline data comparison.