MIT and Sakana AI Cut Cost of Self-Improving Coding Agents with SIFT Framework

Researchers from MIT and Sakana AI have developed a framework called SIFT (Self-Improvement via Fast Tree-search) to reduce the cost of evaluating self-improving coding agents.

Coding agents that rewrite their own code need someone to check whether each tweak improves performance. This can be expensive, especially when an agent generates many candidate modifications.

SIFT sidesteps this problem by using a large language model as a referee to compare two candidate modifications and decide which one looks better. The framework also runs evaluations asynchronously, allowing multiple candidates to be tested at the same time.

The result is a hybrid pipeline where the language model filters out weak candidates cheaply, and only the top performers get the costly downstream evaluation.