Teaching language models to write faster vectorized code

Explicit vectorization using hardware intrinsics remains critical for high-performance computing, but LLMs struggle due to limited training data and strict instruction constraints. AutoVecCoder addresses this with two components: VecPrompt, a data synthesis pipeline that injects domain-specific intrinsic knowledge, and VecRL, reinforcement learning aligned with execution efficiency. The resulting 8B parameter model achieves state-of-the-art results on SimdBench's SSE and AVX subsets, generating code that sometimes surpasses compiler -O3 optimizations. Intended for systems programmers and performance-critical application developers.