Systematically combining agentic exploration with human expertise in kernel generation

Determine a systematic framework for integrating agentic exploration with human expert guidance in GPU kernel generation to expand the design space and improve controllability in performance-critical settings.

Background

The paper highlights human–AI collaboration as a complementary paradigm to fully automated agentic approaches. It emphasizes that, despite rapid progress, a principled method for combining agentic exploration with expert input remains unresolved.

The authors identify two requirements for operationalizing such collaboration: explainability (agents provide interpretable rationales for optimization choices such as tiling) and mixed-initiative interaction (humans specify high-level constraints while agents execute implementation and tuning). These elements aim to balance controllability with scalable automation.