As large language models (LLMs) become more capable, the long‑term value of AI systems lies not only in solving individual requests but also in turning experience and accumulated knowledge into durable, reusable competence. We introduce SimSkill, a self‑evolving agent built around the Simulation of Urban MObility (SUMO) traffic simulator.
SimSkill automatically identifies its capability gaps, generates environment‑grounded tasks, solves them, and verifies solutions through an action‑critic loop. It consolidates experience into three types of memory—episodic, procedural, and semantic—without updating the backbone model. Through autonomous exploration, it constructs a reusable library that spans the entire traffic‑simulation workflow.
We evaluate SimSkill on two held‑out benchmarks using three different backbone LLMs and independent artifact‑based verification. Results show up to a 25‑percentage‑point increase in verified completion. Ablation studies reveal complementary contributions from procedural and semantic memory. The benefits of memory depend on the model and budget: not every model gains from memory, nor does it uniformly reduce inference cost.
More broadly, SimSkill illustrates a design paradigm where natural language preserves and composes computational capabilities, while executable tools and code provide precise and reproducible execution. All code and experimental data are publicly available at https://github.com/qiliuchn/SimSkill-V1.
Review