On-Device Vision-Language Models Through Link Outages

Under review at ACM MobiCom 2027

Mobile edge systems such as drones and robots increasingly rely on cloud-hosted vision-language models for visual reasoning, and lose that capability whenever the wireless link drops. Falling back to a local model is not free: loading it only after the link fails wastes much of the outage, while keeping it resident all the time takes memory and compute away from the workloads the device is actually there to run.

This project uses forecasts of upcoming connectivity loss to prepare a suitably sized local model during the warning interval, so it is ready when the outage begins and released once the link recovers. It also carries useful context from the cloud model across the disconnection and serves queries within their deadlines during the outage. The system is implemented and evaluated on an embedded edge platform with real wireless traces and aerial-imagery workloads.

Details will be shared after the review process.