Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching
Aaron Havens, Benjamin Kurt Miller, Bing Yan, Carles Domingo-Enrich, Anuroop Sriram, Brandon Wood, Daniel Levine, Bin Hu, Brandon Amos, Brian Karrer, Xiang Fu, Guan-Horng Liu, Ricky T. Q. Chen · 2025
Read Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching by Aaron Havens, Benjamin Kurt Miller, Bing Yan, Carles Domingo-Enrich, Anuroop Sriram, Brandon Wood, Daniel Levine, Bin Hu, Brandon Amos, Brian Karrer, Xiang Fu, Guan-Horng Liu, Ricky T. Q. Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-11713.txt.
Description
Aaron Havens and co-authors wrote a 2025 paper on a faster way to train diffusion samplers. Such samplers learn to draw samples from a target defined only by an energy function, a common task in physics and chemistry. Past methods needed many costly energy checks, which capped the size of problems they could handle.
Their method, Adjoint Sampling, rests on stochastic optimal control theory. It allows far more training updates than energy checks, which makes training much cheaper. It also keeps the guarantees of an earlier method without needing fixes that push samples toward the target. Tests cover classic energy functions and neural network energy models. The team even generates molecule shapes across many systems with one trained model. They plan to release the hard test cases openly.
Catalog record
| CREATOR | Aaron Havens, Benjamin Kurt Miller, Bing Yan, Carles Domingo-Enrich, Anuroop Sriram, Brandon Wood, Daniel Levine, Bin Hu, Brandon Amos, Brian Karrer, Xiang Fu, Guan-Horng Liu, Ricky T. Q. Chen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 119 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-11713