Researchers proposed a self-improvement framework for coding agents. It uses a large language model as a judge to compare candidate patches. Win-loss records are aggregated with a Bradley-Terry model. Strength scores guide a lightweight tree search. Downstream evaluations are reserved for the most promising candidates. The method is called Recursive Self Improvement via Fast Tree-search. It outperforms prior tree-search frameworks on Polyglot. It requires far fewer CPU hours, wall clock time and API costs.