AgentCompile: An LLM-Guided Compiler for Direct CUDA Inference

This is an arXiv feed of recent programming-languages / compiler / systems research papers (June 9, 2026) covering: AgentCompile — an LLM-guided CUDA inference compiler that uses LLMs as advisory metadata to pick and empirically validate CUDA templates (reports multi-fold inference speedups; will be open-sourced); SNN-MLIR — an out-of-tree MLIR dialect and NIR→MLIR→C toolchain for spiking neural networks that emits dependency-free C11 for CPUs/embedded targets (Apache-2.0); LongRTL — an LLM+graph-similarity framework for long-context RTL decomposition, optimization, and reconstruction; a world