MIT and Sakana AI debut SIFT to slash coding-agent evaluation costs

Researchers at MIT and Sakana AI have introduced Recursive Self-Improvement via Fast Tree Search (SIFT), a framework designed to bypass the primary bottleneck in self-improving coding agents: the massive compute cost of running full benchmarks. Instead of executing every candidate modification against heavy test suites, SIFT employs a separate language-model judge to pairwise-compare code changes asynchronously. In tests on the Polyglot coding benchmark, a SIFT-managed agent reached 35.1% accuracy in under five hours of wall-clock time while consuming roughly $150 in API credits.

VentureBeat · October 2, 2026 · AI & Frontier