A Unified Density Operator View of Flow Control and Merging
Riccardo De Santi, Malte Franke, Ya-Ping Hsieh, Andreas Krause · 2026
Read A Unified Density Operator View of Flow Control and Merging by Riccardo De Santi, Malte Franke, Ya-Ping Hsieh, Andreas Krause (2026) 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-2602-08012.txt.
Description
Flow and diffusion models face two separate challenges: steering a pretrained model with rewards, and merging several models into one. Riccardo De Santi and three co-authors unify both problems in a single probability-space framework built from operators over model densities.
Within it, intersection can enforce safety, union can compose diverse models, interpolation can aid discovery, and all of these gain reward-guided versions, combinable into logical expressions via generative circuits. Their reward-guided flow merging method reduces the task to a sequence of standard fine-tuning problems solved by mirror descent. Generative-model researchers gain a principled language for combining and controlling models instead of ad hoc tricks.
Catalog record
| CREATOR | Riccardo De Santi, Malte Franke, Ya-Ping Hsieh, Andreas Krause |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-08012