Automatically generating high-performance kernels with LLM-based agents could greatly improve development efficiency, according to a new arXiv paper. The authors identify two remaining obstacles, the first being the difficulty of increasing multi-turn interaction with execution environments—though the abstract leaves the second obstacle unspecified.
The paper proposes a compaction-aware reinforcement learning approach for Ascend Triton kernel generation. By incorporating compaction awareness into the RL loop, the method aims to overcome the challenges that currently limit LLM-driven kernel generation.
As the source is a single abstract, details on the method's implementation and results are not yet available. The work appears to target Ascend-specific hardware, suggesting a focus on practical deployment in that ecosystem.