SkVM: Language Virtual Machine Makes LLM Agent Skills Portable Across Models and Harnesses
arXiv·medium signal
Researchers at SJTU-IPADS published SkVM (arXiv 2604.03088), a compilation and runtime system that treats LLM skills as code and models as heterogeneous processors. It decomposes skill requirements into primitive capabilities, profiles model-harness pairs, and performs AOT compilation with JIT optimization at runtime. Results show significantly improved task completion rates across different models while reducing token consumption by up to 40%. Open-sourced on GitHub — directly relevant to anyone building multi-model agent workflows.