KV-Skill: Forging Expertise in the Model's Native Language

The paper addresses the trade-off between text prompts, which are modular but require interpretation on every use, and weight updates, which are effective but difficult to load, remove, or share independently. KV-Skill introduces a design space of external factorized operators that a frozen language model accesses through a lightweight interface, keeping the base model unchanged while enabling modular capability injection. The abstract mentions two complementary paths, with 'Registration' as one; the second is not fully described in the excerpt. This approach has potential implications for model deployment, enabling plug-and-play swapping of skills and more flexible reuse of expertise across applications. The preprint is numbered 2608.05475v1 and appears in the cs.LG category, indicating a machine learning methodology focus.