A team led by Professor Kyunghan Lee of Seoul National University presented two developments intended to ease the operation of artificial intelligence (AI) on smartphones and other devices. The first, TACToC, reduces the volume of transmitted data; the second, mzCache, speeds up AI response after switching between applications. Both works were accepted to the ACM MobiCom 2026 conference, which will be held in Austin (Texas) from October 26 to 30.
Modern video compression systems were created for the human eye, so they transmit a lot of information that is unnecessary for machine analysis. TACToC is a task-oriented communication system: it transmits only the visual features that AI needs for a specific goal. Unlike previous approaches that required separate training for each task, this uses a single encoder, and the task is specified by a small "prompt" - a set of parameters. When a new task is added, only these parameters are trained, not the entire model.
In experiments on five tasks (semantic segmentation, pose estimation, image captioning, classification and depth estimation), TACToC reduced data transmission costs by 1.5-14.5 times compared with existing methods, while the drop in accuracy did not exceed 3 percentage points. Switching between tasks required updating about 1,900 parameters and took approximately 12.5 milliseconds. On an NVIDIA Jetson AGX Orin device, the system processed about 21 frames per second. In a real mobile network environment, TACToC reduced bandwidth usage by 16.7% compared with the H.265 codec and by 41.1% compared with another method.
The second development, mzCache, addresses the problem of insufficient RAM. When a user switches from an AI application to the camera or a game, the operating system frees memory, and on returning to the AI one has to wait for data to be restored. Measurements on smartphones showed that the time to first response increased from 0.8 seconds to about 16 seconds. mzCache divides AI memory into small blocks and, when the user returns, restores data in parallel with computation: the graphics processing unit (GPU) computes on already available data while the central processing unit (CPU) loads the rest. In experiments on the Galaxy S25+ and OnePlus 12, the system sped up the first response by 2.1-5.5 times compared with the conventional method, and with full memory release by up to 25.9 times compared with the standard Android approach. In addition, mzCache reduced energy consumption for restoration and computation.
Both technologies are currently research prototypes. TACToC does not require changes to existing network architectures, and mzCache is implemented as an Android application that does not affect the operating system kernel. The scientists plan to extend the developments to environments with multiple AI applications and to a broader range of mobile processors, including neural processing units (NPUs).