AdaExplore Framework Enables LLM Self-Improvement for Kernel Code Generation

AdaExplore is a novel framework designed to overcome the shortcomings of large language model agents in generating performance-critical kernel code. Existing methods often handle each problem instance in isolation, failing to build on reusable knowledge, which is especially challenging for domain-specific languages like Triton that lack representation in LLM pretraining datasets. These languages impose strict constraints and feature complex optimization landscapes, making straightforward generation and local refinement unreliable. AdaExplore facilitates self-improvement through accumulated execution feedback in two phases: failure-driven adaptation and diversity-preserving search. This combined strategy enhances both correctness and optimization performance without needing extra fine-tuning or external information. The framework was outlined in a research paper on arXiv, identified as 2604.16625v1, and focuses on performance-critical kernel code generation, an area where robust self-improvement has yet to be resolved.