MIT AND SAKANA AI MAKE SELF-IMPROVING AGENTS CHEAPER TO EVALUATE

Summary from the TLDR IT newsletter

SIFT uses an LLM judge to rank proposed coding-agent improvements before spending compute on full benchmark runs. One experiment reached 35.1% accuracy while using substantially less compute than previous search approaches.