We propose a novel self-improvement algorithm to teach language models to perform effective search.