Zheng Ren

Language models learn search strategies via practice, transferable as text

Researchers have developed a method called Code-to-Harness that enables language models to learn numerical search strategies through practice and then distill these strategies into text. This approach significantly redu…