SkVM: A Language VM for Skill Enables One‑Write, Everywhere‑Efficient Execution on Any LLM

Skill–Model Mismatch

Empirical analysis of 118 k Skills shows that 15 % of tasks suffer performance degradation when a Skill is added, 87 % of tasks have at least one model with no improvement, and some Skills increase token consumption by up to 451 % while success rates remain unchanged.

Skill is written as “natural‑language code,” but the semantic gap between Skill requirements and the heterogeneous capabilities of LLMs and Agent harnesses leads to inefficiency.

SkVM Overview

SkVM is a language virtual machine designed for Skill, inspired by classic VMs such as the Java Virtual Machine. It enables a single Skill implementation to run efficiently on any model and Agent harness.

PASS‑1: Ability‑Based Compilation

SkVM extracts 26 primitive capabilities from the target LLM to form a capability profile. It analyses each Skill to determine required capabilities and their levels. If a Skill’s required level exceeds the model’s provision, the compiler rewrites the Skill to lower its demand.

Example: when a model lacks path‑resolution capability, relative file paths in Skill scripts are converted to absolute paths, reducing dependence on the “script execution” capability.

PASS‑2: Environment Binding

Skills often declare required packages. SkVM’s ahead‑of‑time (AOT) compiler automatically extracts these dependencies, generates installation and verification scripts, and binds the environment before execution, eliminating token‑wasting trial‑and‑error by the LLM.

PASS‑3: Concurrency Extraction

More than 76 % of Skills contain workflows that are executed serially by default.

SkVM discovers data‑parallel, instruction‑parallel, and thread‑parallel opportunities, constructing a DAG workflow that can be executed concurrently.

Developers can register custom optimization passes to further improve parallelism.

Runtime Optimizations

JIT and Code Solidification

During execution, SkVM matches generated code fingerprints against previously compiled templates. When matches succeed repeatedly, SkVM applies JIT compilation to solidify the executable code, avoiding repeated LLM generation of scripts and saving tokens.

Adaptive Re‑compilation

If runtime errors occur, SkVM feeds the logs back to the compiler, which automatically re‑optimizes the Skill to prevent recurring failures and improve success rates.

Resource‑Aware Parallelism

SkVM adjusts parallel granularity based on current system resources, reducing contention and further improving throughput.

Experimental Results

The team evaluated SkVM on 118 representative tasks covering code generation and data analysis. Key findings:

Small models (e.g., Qwen‑30B) achieve task success rates comparable to Opus 4.6 after SkVM compilation.

Top‑tier models see token consumption reduced by up to 40 %.

Execution latency for code‑heavy steps drops from tens of seconds to a few hundred milliseconds, a 19‑ to 50‑fold speedup.

Parallelism extracted by SkVM yields up to 3.2× efficiency gains.

SkVM integrates seamlessly with major Agent frameworks such as OpenClaw, Hermes, OpenJiuwen, and PI, and supports the Clawhub Skill ecosystem.

Paper: https://arxiv.org/abs/2604.03088

Project site: https://skillvm.ai/

Repository: https://github.com/SJTU-IPADS/SkVM/