A new framework lets coding agents rewrite their own code and pick the best version without burning through a testing budget to find it.
Researchers describe SIFT (Self Improvement via Fast Tree-search), a system for agents that recursively modify their own implementations. Earlier self-improvement setups judged each candidate patch by rerunning it against a slice of benchmark tasks - accurate, but slow enough to become the whole bottleneck. SIFT adds a cheaper filtering step first: an LLM acts as judge, comparing patches head-to-head, and those pairwise results feed a regularized Bradley-Terry model, the same style of math used to rank players from win-loss records. Only the patches that come out on top get the expensive full evaluation, run on the Polyglot coding benchmark inside a disaggregated tree search.
The real constraint on self-improving code agents has never been whether the idea works - it has been paying for enough evaluation runs to tell good patches from bad ones. Reserving the costly benchmark runs for a small shortlist, instead of every candidate, is what lets SIFT reportedly beat prior tree-search self-evolution approaches on the full Polyglot benchmark while using less CPU time, less wall-clock time, and less API spend.
Worth remembering: the judge here is still a model guessing which patch is better, not a ground-truth signal. If it misjudges early, the search can prune away the strongest candidates before they ever reach a real test.