KLineage: Learning Optimization Timing from Expert GPU Kernels

A new arXiv preprint (2605.28213) introduces KLineage, a system that teaches LLM-based agents when to apply GPU kernel optimizations by learning from expert implementations. Instead of forward rollouts, KLineage walks expert kernels backward through validation-gated simplifications, reversing each accepted step into a reusable skill. Each skill records the optimization intent, code location, validity conditions, effects, and potential failure modes. A downstream LLM applies these skills to new code under compile/correctness/profile gates. On five expert workloads across two NVIDIA architectures, lineage-derived skills outperformed recent memory-based LLM-kernel baselines in final kernel quality.