SILO: Generative Modeling Without the Target Data: Inference-Time Steering and Distribution Inversion

Abstract

Transport-based models such as score-based diffusion and flow matching have become a leading paradigm for generative modeling: given samples from a distribution, one learns dynamics that produce new ones. Many scientific problems do not come in this form. The distribution of interest may be a tilted version of one we can already sample, or one we never observe directly and see only through indirect, noisy measurements.

I will describe two directions addressing these gaps. First, inference-time steering: reaching a new target from a pre-trained model without retraining. Guidance is simple but biased, while particle-based corrections such as Sequential Monte Carlo suffer from weight degeneracy. We introduce DriftLite, a lightweight, training-free method exploiting a previously unexplored degree of freedom in the Fokker–Planck equation between the drift and the particle potential, demonstrated on large-scale protein–ligand co-folding. Second, learning a generative model when clean data never existed and the forward process is available only as a black-box simulator. This is common in the structural sciences, where every recorded structure was inferred rather than measured. The Self-Consistent Stochastic Interpolant pins down the transport map by requiring that its composition with the forward model reproduce the observed corrupted distribution. The resulting scheme needs only forward evaluations, admits convergence guarantees, and performs strongly on imaging and scientific reconstruction.

Based on joint work in arXiv:2509.21655 and arXiv:2512.10857

Bio

Jiequn Han is a Research Scientist at the Center for Computational Mathematics, Flatiron Institute, Simons Foundation. His research develops machine learning methods for high-dimensional problems in scientific computing, with current work on generative modeling, sampling, and inverse problems, building on a longer-standing line on high-dimensional PDEs and multiscale modeling of physical systems. He holds a Ph.D. in Applied Mathematics from Princeton University and dual bachelor’s degrees in Computational Mathematics and Economics from Peking University. His research has been recognized with the SIAM Computational Science and Engineering (CSE) Early Career Prize.