SIFT Enables Budget-Efficient Self-Improvement for Coding Agents

Summary

Coding agents can improve their own implementations by recursively proposing and testing self-modifications, but existing methods require costly repeated benchmark evaluations. The paper introduces Recursive Self Improvement via Fast Tree-search (SIFT), a sample-efficient framework designed for strict compute and API budgets. SIFT evaluates candidate patches with two signals: expensive downstream coding-task tests and faster pairwise comparisons produced by an LLM judge. The pairwise win-loss results are aggregated with a regularized Bradley-Terry model to estimate candidate strength. Those scores guide rank-based parent selection in a lightweight, fully disaggregated tree search, while downstream task evaluations are reserved for the most promising nodes. This intermediate signal lets the search explore promising patches without waiting for every candidate to undergo slow benchmark testing. On the full Polyglot benchmark, the authors report that SIFT outperforms existing tree-search-based self-evolution frameworks while using fewer CPU hours, less wall-clock time, and lower API cost.