AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation

Create an agentic framework for specialized DSL (like Triton) that uses failure-driven adaptation and stores reusable kernel knowledge. This addresses the bottleneck where current LLM coding agents fail to learn from repeated kernel optimization attempts.

Suggested repo: triton-refiner

"AI agents that actually learn to write faster kernels."

Estimated effort: 120h