What's this about?
Prompt optimization is a promising route to improving agents across diverse tasks, and is central to arguments about recursive self-improvement. This paper questions whether the elaborate search methods that have accumulated around it are needed at all, and finds a naive approach competitive.
What is "recursive self-improvement"?
A system that improves itself, so that the improved version improves itself further, and so on. It is the mechanism behind both the optimistic and catastrophic long-run scenarios, and prompt optimization is the most tractable place to study it — a model rewriting its own instructions is self-improvement you can run on a laptop, with no weights touched.
Why does this matter to you?
- "Does the simple version work as well" is a question worth asking of every optimization method, and the answer is embarrassingly often yes.
- If naive optimization is competitive, then RSI experiments are cheaper and more accessible than assumed — which cuts both ways.
- Same genre as SWE-Prime and the SFT-versus-RL paper from the 2026-08-28 digest: the field is auditing its own complexity.