The artificial intelligence on your smartphone could become noticeably faster without relying so heavily on cloud servers. Researchers at KAIST have built a system called CURE, or Cumulative Knowledge Reuse, that lets a small on-device AI model remember and reuse answers from a larger server model. The approach cut server calls by 55.61% while keeping accuracy close to that of sending everything to the cloud.
How CURE saves your phone a trip to the server
Phones typically run small, lightweight AI models due to limited processing power and memory. These models handle simple tasks fast but struggle with complex or unfamiliar images, so they lean on a server model for help.
Older systems treat each server answer as a one-time fix, and your smartphone asks again when a similar situation comes up. The shift toward local AI is already well underway, with phones increasingly handling tasks directly on-device.
CURE changes this process with a three-step check. It first sees if the device can solve the task alone. Then it checks stored knowledge from past server help, and only then contacts the server. Say your phone can’t identify a car model in a photo. The server answers once, and CURE stores the car’s key features instead of the full image. Your phone can then recognize similar cars on its own the next time you ask.
What this means for future AI devices
In testing, researchers paired CURE with vision-language models using MobileCLIP2 on the device with the far larger, 18 billion parameter EVA-CLIP model on the server.
End-to-end testing, including communication time, showed real speed gains, with CURE running up to 2.80 times faster than conventional device-server collaboration. It was up to 3.67 times faster than sending everything to the server.
CURE requires no retraining and only adds a separate knowledge store, so this could roll out without replacing your phone’s AI. That could become increasingly important as smartphones move toward an AI-first era. The approach could also help robots and wearables, especially with limited computing power and unreliable connections.
It goes without saying that the gains will depend on device hardware, storage, networks, and input data. Further testing is still needed, but CURE offers a simpler path toward faster, more efficient on-device AI.
