Researchers from MIT and Sakana AI have introduced SIFT (Self-Improvement via Fast Tree-search), a framework designed to reduce the computational costs of verifying self-improving coding agents. The method uses a large language model to compare candidate code modifications head-to-head, rather than running full benchmarks for every change.
The framework utilizes a regularized Bradley-Terry model to rank preferences and performs evaluations asynchronously. This hybrid pipeline filters candidates before applying costly downstream testing. The paper, authored by Xinghong Fu, Aravinth Kulanthaivelu, and Yutaro Yamada, was published on arXiv on September 18, 2026.
In testing with an o3-mini agent on the Polyglot benchmark, SIFT achieved a score of 35.1% in 30 expansion steps, outperforming the Darwin Gödel Machine (DGM) which reached 30.7% in 80 nodes. Using a Qwen3-Coder-30B configuration, the search required 224 CPU hours and approximately $34 in API costs, roughly one-tenth of the resources used by DGM.
SIFT also demonstrated performance improvements on other benchmarks. On TerminalBench 2.1, scores rose from 29.2% to 36.7%, while on SWE-60, the results increased from 40.0% to 52.1%. Discussions surrounding the research highlight both the potential for increased accessibility and the critical reliance on the quality of the language model judge.
