MIT and Sakana unveil SIFT to cut self improving coding agent costs

SIFT matters because it attacks one of the least glamorous but most real bottlenecks in agent research: the cost of evaluation. Self improving coding agents can generate endless variants of themselves, but every variant has to be tested somewhere. When each full benchmark run costs serious money in API calls and compute time, only a handful of well funded labs can really explore large search trees. By inserting a language model judge that does cheap pairwise comparisons and only sends the most promising candidates to full evaluation, SIFT turns that bottleneck into something closer to a scheduling problem.

If the results hold up, this could dramatically broaden who can run serious self improvement experiments. Ranzware’s summary highlights a Polyglot score of 35.1 percent for about $150 in API and 42 CPU hours in one configuration and even cheaper runs using open weights like Qwen3 Coder. Those numbers move recursive self improvement from a toy demo for o3 class proprietary models into something smaller labs can tinker with on modest budgets.

The flip side is that this pushes more leverage onto the judging model itself. If your LLM referee consistently prefers flashy but brittle changes, you can steer an agent in the wrong direction cheaply. That aligns with a deeper trend in the race to AGI, where meta systems that select, route and evaluate models become as important as the base models. SIFT is an early, concrete example of that shift.