Tag
This paper introduces CUDAnalyst, a tool for analyzing how individual feedback signals influence planning decisions in self-evolving LLM agents for CUDA kernel generation, using trajectory freezing and selective feedback injection to enable controlled attribution.