KAIST AI Cuts Smartphone Server Calls by 55 Percent

Researchers at KAIST have introduced a method that allows small AI models on smartphones to retain and reuse knowledge previously provided by powerful server-based systems. This approach enables mobile devices to handle complex tasks locally without needing to consult the cloud for every single instance, significantly reducing the time users wait for responses.

The technology, known as CURE, acts like a digital notebook where the phone records the solutions it receives from the server. By storing these insights, the device can apply them to similar future problems independently. According to reports from techxplore.com, this cumulative reuse strategy reduced the number of required server calls by an average of 55.61% while maintaining high accuracy levels.

Reducing dependence on cloud processing

Smartphones operate with limited memory and processing power, which restricts the complexity of the AI models they can run directly. Typically, when a local model encounters a difficult image or task, it sends the data to a large server model for a definitive answer. While this ensures accuracy, it introduces network delays and increases computational costs for service providers.

Previous collaborative approaches treated each server response as a one-time event. If the phone faced a similar problem later, it had to request help again, repeating the same cycle of data transfer and waiting. CURE breaks this pattern by converting those one-time answers into persistent knowledge that the device can access offline, allowing for faster decision-making on repeat or similar tasks.

Storing abstract features instead of images

A key aspect of CURE is how it stores information. Rather than saving copies of the original photos or data processed by the server, the system records a summary of shared features and differences. This is comparable to noting specific distinguishing characteristics of an object rather than memorizing the entire visual input.

For example, if a server helps identify a specific car model, CURE stores the relevant visual cues like body shape or headlight design. This abstracted knowledge allows the phone to recognize not only the exact image it saw before but also new, similar images it encounters for the first time. This method ensures the device remains efficient without bloating its storage with redundant raw data.

Balancing accuracy with efficiency gains

The research team tested CURE using vision-language models that connect visual data with text. They used a lightweight model on the device and a much larger model with 18 billion parameters on the server. The results showed that CURE achieved accuracy levels close to those of a system that sends every input to the server.

The trade-off is minimal. While the system relies on initial server interactions to build its knowledge base, it quickly becomes more self-sufficient. This reduces the overall network traffic and computational load on servers, offering a practical path toward faster, more efficient mobile AI without sacrificing the precision required for complex tasks.