Goal-Persistent Coding Agents Optimize Scientific GPU Software

Summary

This arXiv paper examines whether general-purpose coding agents can perform rigorous scientific performance engineering over many tool-use turns. The authors give off-the-shelf Codex and Claude Code an executable goal for optimizing fixed-radius nearest-neighbor search used in particle tracking, including exact-correctness tests, profiling requirements, and acceptance criteria, while leaving the code transformations unspecified. Starting from a PyTorch-dependent CUDA implementation, the primary sequential run removed the PyTorch dependency and used hypothesis-driven optimization experiments to produce a standalone C++/CUDA library. The library exactly reproduced the targeted reference result. Its synchronous NumPy interface was 1.6 times faster than the original GPU-resident PyTorch interface, even after including host transfers. Comparable speedups appeared across different GPU architectures and software stacks. An independent rerun followed a different sequence of hypotheses and achieved still better performance on the target workload. The authors argue that persistent coding agents can act as experimental performance engineers, but that executable scientific contracts are needed both to direct optimization and to validate its results.