Researchers at Seoul National University have introduced two technical systems designed to address data traffic and memory bottlenecks in artificial intelligence applications. Presented at the ACM MobiCom 2026 conference in Austin, Texas, the technologies aim to improve performance for mobile devices, autonomous vehicles, and intelligent monitoring systems.
Reducing Data Overhead with TACToC
The TACToC (Task-Adaptive Communication for Task-oriented Communications) system addresses the inefficiency of transmitting full video streams for AI analysis. Conventional methods often send unnecessary visual data that exceeds the requirements of specific AI tasks. Tech Xplore reported that TACToC utilizes a single "task-describable encoder" that selectively transmits only the visual features essential for a given task, such as object recognition or depth estimation.
By using trainable tokens to define tasks, the system avoids the need to retrain encoders for every new function. In experiments using five public datasets—including semantic segmentation and image captioning—TACToC reduced communication costs by 1.5 to 14.5 times compared to existing methods. The system maintained accuracy within a 3-percentage-point margin. According to the research team, TACToC processed approximately 21 frames per second on an NVIDIA Jetson AGX Orin edge AI device, offering a potential path for future 6G semantic communications.
Addressing on-Device Memory Bottlenecks via mzCache
Smartphones often struggle to balance AI model requirements with the demands of other background applications. When an operating system reclaims memory from an AI model to support a new app, the AI’s "time to first response" can spike significantly. Researchers observed that on-device AI response times on standard smartphones increased from 0.8 seconds to approximately 16 seconds following memory reclamation.
The team’s solution, mzCache, manages memory by dividing AI data into smaller units. It allows the system to recover only the necessary memory segments while simultaneously running AI computations. By prioritizing the restoration of earlier AI model layers, mzCache reduced response delays by 2.1 to 5.5 times compared to conventional storage-based recovery. In testing, the system allowed AI assistants to remain active and preserve conversation context even when users cycled through 10 different apps over a 45-minute period.
Future Deployment in Mobile Environments
Professor Kyunghan Lee, the corresponding author for both papers, noted that these developments provide a foundation for stable AI services across diverse hardware, including smart glasses and mobile processors. The research team implemented mzCache in approximately 6,000 lines of code, designed to function as an Android application without requiring modifications to the operating system kernel.
Future development will focus on scaling these technologies to support environments where multiple AI applications run concurrently. The researchers also plan to broaden the compatibility of these systems to include a wider range of hardware, such as neural processing units (NPUs).
Technical Implementation Details
- TACToC Performance: Achieved 1.5× to 14.5× reduction in communication costs; maintained accuracy within 3 percentage points.
- mzCache Efficiency: Reduced time to first response by up to 5.5× compared to conventional storage recovery; performed up to 25.9× faster than default Android memory management when memory was fully reclaimed.
- Research Presentation: Both papers were accepted for the 32nd Annual International Conference on Mobile Computing and Networking, held in Austin, Texas, from Oct. 26 to 30, 2026.
Frequently Asked Questions
Does TACToC require new hardware to function?
No, TACToC is designed for deployment within existing communication network architectures. It is intended to be compatible with current mobile and edge AI hardware, such as the NVIDIA Jetson AGX Orin.
How does mzCache prevent AI termination during multitasking?
Instead of forcing the AI to offload its entire state to storage, mzCache divides the memory into smaller units. It only reclaims the specific memory needed by the OS, allowing the AI to keep its conversation context active in the background.
Can mzCache be used on any smartphone?
The researchers tested mzCache on commercially available devices like the Galaxy S25+ and OnePlus 12. Because it runs as an Android application, it does not require kernel-level modifications, suggesting potential for broader mobile device support.