AI Regex Scientist: A self-improving regex solver

2 days ago

I built a system where two LLM agents co-evolve: one invents regex problems, the other learns to solve them. The generator analyzes the solver's failures to create challenges at the edge of its abilities.

The result: autonomous discovery of a curriculum from simple patterns to complex regex, with a quality-diversity archive ensuring broad exploration.

Blog: https://pranoy-panda.github.io/2025/07/30/3rd.html

Code: https://github.com/pranoy-panda/open-ended-discovery

2 comments

PranoyP

Reply
allinonetools_  8 hours ago

This is a nice example of curriculum emerging from feedback rather than being hand-designed. In my experience, systems improve much faster when the task generator adapts to actual failure modes instead of fixed difficulty levels. The quality-diversity angle is especially interesting here.