The paper introduces P2Skill, a method for privacy-preserving skill distillation in cloud-local LLM inference systems. The setup is common: a small model runs locally, but for harder tasks it can call on a larger cloud model. The catch is that any request sent to the cloud must be stripped of personally identifiable information (PII).

The authors note that existing privacy-preserving techniques work by modifying the prompt or the request in some way. P2Skill instead appears to focus on distilling skills from the cloud model into the local model, so that the local model can handle more tasks on its own, reducing the need to send sensitive data outward.

Because the abstract is truncated, the specific mechanism and evaluation results are not available in the source. The paper is a single source, so there are no conflicting findings to compare. Its contribution is positioned as a new direction within the broader effort to make cloud-assisted LLM inference both capable and privacy-compliant.