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

TL;DR

AdaExplore is a novel agent framework that uses failure-driven adaptation and diversity-preserving search to improve kernel code generation by accumulating execution feedback, enhancing correctness and performance without extra fine-tuning.

Contribution

It introduces a two-stage approach enabling self-improvement through feedback, memory, and structural search, specifically for domain-specific language kernel optimization.

Findings

Achieves 3.12x speedup on KernelBench Level-2

Achieves 1.72x speedup on KernelBench Level-3

Continues to improve with more computation

Abstract

Recent large language model (LLM) agents have shown promise in using execution feedback for test-time adaptation. However, robust self-improvement remains far from solved: most approaches still treat each problem instance independently, without accumulating reusable knowledge. This limitation is particularly pronounced in domain-specific languages such as Triton, which are underrepresented in LLM pretraining data. Their strict constraints and non-linear optimization landscape further make naive generation and local refinement unreliable. We propose AdaExplore, an agent framework that enables self-improvement via accumulated execution feedback for performance-critical kernel code generation through two complementary stages: failure-driven adaptation and diversity-preserving search, jointly improving correctness and optimization performance without additional fine-tuning or external knowledge. In the adaptation stage, the agent synthesizes tasks and converts recurring failures into a reusable memory of validity rules, helping subsequent generations remain within the feasible set. In the search stage, the agent organizes candidate kernels as a tree and alternates between small local refinements and larger structural regeneration, allowing it to explore the optimization landscape beyond local optima. Experiments on kernel runtime optimization benchmarks validate these gains: AdaExplore achieves 3.12x and 1.72x speedups on KernelBench Level-2 and Level-3, respectively, within 100 steps, and continues to improve with additional computation.