Bilevel autoresearch loops — an outer coach loop rewrites the search

Most people don't realize that you can get 5x more efficiency from a loop system by just adding another loop on top. Here's how the bi-level loop system works. If you don't already, Andre Carpethy, one of the co-founders of OpenAI, released a loop engineering system that works when you give AI a goal and it works towards it by itself. It tries to change and test whether the results get better. Keeps it if it did, throws a weight if it didn't, and tries the next idea all automated. But here's a problem nobody noticed. Even good loops have bad habits. AI has instincts about what kinds of ideas to try. It keeps going back over the same instincts over and over even after they stop working. Your loop runs 500 experience, but it's really running the same 10 ideas 50 times. This is a fix that Frontier Lab scientists found out. You just add a second loop on top of the first one. The inner loop is the worker. It keeps running experiments exactly like before while the top loop is the coach. The coach does not do any of the work. Its only job is to watch the worker. The coach in the outer loop rewrites the worker's playbook, changing the way it searches for ideas and then setting it back in. Researchers tested this on the same exact benchmark that Carpethy used, and this new Tula version scored five times better than a single loop with the same exact LLM models. The entire improvement came from the structure that one AI is forcing the other AI to explore ideas that normally its…