AI
MIT and Sakana AI cut the cost of self-improving coding agents
Researchers at MIT and Sakana AI released SIFT, a framework that uses a separate language model as a pairwise judge to rank candidate rewrites of a coding agent before expensive benchmark runs are spent on them. The paper reports that SIFT beats existing tree-search self-evolution methods on the multilingual Polyglot benchmark while using significantly less CPU time, wall-clock time and API spend. It builds on the Darwin Godel Machine harness and is described in the paper 'Self Improvement via Fast Tree-search'.