<figcaption style="text-align: center;">Overview of the problem that we study and the general structure of its solutions. Given a collection of different code sources, where a source is either program or repository, we design a modular and reusable library. To do this we cluster together related programs into tuples (left), sample different rewrites using a language model (right), and select the rewrites which optimize various criteria, such as compression, while validating candidate rewrites using test cases (pass rate).</figcaption>
0 commit comments