Coding agents can recursively modify their own implementations, creating a self‑improvement loop. Prior work shows this can boost performance on coding benchmarks, but existing methods are costly and compute‑intensive. We propose a simple, sample‑efficient self‑improvement framework that markedly improves coding performance under strict budget constraints. We identify the evaluation of candidate self‑modifications as the main runtime bottleneck because traditional approaches re‑run a subset of benchmark tasks with the modified agent, which is time‑consuming. To address this, we introduce Fast Tree‑Search Self‑Improvement (SIFT), which augments downstream task evaluations with an LLM‑as‑a‑judge signal that performs pairwise comparisons of candidate patches. The win‑loss records are aggregated using a regularized Bradley‑Terry model, and the resulting strength scores drive rank‑based parent sampling in a lightweight disaggregated tree search. Expensive downstream evaluations are reserved for only the most promising nodes. The fully disaggregated tree‑search pipeline lets judge scores provide intermediate guidance during exploration, avoiding slowdown from slow evaluation runs. Experiments on the Polyglot benchmark show that SIFT outperforms existing tree‑search‑based self‑evolution frameworks while requiring far fewer CPU hours, wall‑clock time, and API costs.
Review