SIFT — Recursive Self-Improvement via Fast Tree-search (arXiv:2609.19526v1)
The paper introduces SIFT, a sample-efficient self-improvement framework for coding agents that uses an LLM-as-a-judge to perform pairwise comparisons of candidate code patches. Judge scores are aggregated with a regularized Bradley–Terry model to guide a lightweight disaggregated tree search, reserving expensive downstream task evaluations for the most promising nodes; the method reportedly outperforms prior tree-search self-evolution approaches on the full Polyglot benchmark while using far fewer CPU hours, wall-clock time, and API cost.