SNU Professor Kyunghan Lee’s team has two papers on AI communication and memory technologies accepted to MobiCom
Develops “TACToC,” which transmits only the information needed for a given task, reducing communication costs by up to 14.5× / Develops “mzCache,” which reduces AI time to first response by up to 5.5× when users return to an AI app
image:
Conventional task-oriented communication requires a separate encoder and decoder to be trained for each task, whereas TACToC uses a single encoder to select and transmit only the visual features required for a given task. New tasks can be added by training small task tokens.
view moreCredit: © 2026 Copyright held by the owner/author(s). Huijeong Choe et al., originally published in the Proceedings of ACM MobiCom 2026. Licensed under CC BY 4.0. https://doi.org/10.1145/3795866.3796689
Researchers have developed technologies that reduce communication overhead by transmitting only the data needed for artificial intelligence (AI) analysis and shorten the time users must wait for a response when returning to an AI application after using other apps on a smartphone.
A research team led by Professor Kyunghan Lee of the Department of Electrical and Computer Engineering at Seoul National University College of Engineering has developed TACToC, a communication system that selectively transmits only the information required for AI tasks, and mzCache, a memory management system that reduces response delays for smartphone AI.
TACToC reduced communication costs by up to 14.5× compared with existing methods while keeping the loss in accuracy within three percentage points. In environments where multiple apps are used, mzCache reduced the AI’s time to first response by up to 5.5× compared with conventional storage-based recovery methods.
The two papers describing the research were accepted to ACM MobiCom 2026, a leading international conference in mobile computing and networking, held in Austin, Texas, from October 26 to 30.
“TACToC” Selectively Transmits Only the Information Needed for AI Tasks
Inference offloading, in which video captured by smartphones, autonomous vehicles, or robots is sent to a remote server for AI analysis, can reduce the computational burden on devices. Conventional video compression, however, is designed around visual quality for human viewers and therefore transmits information that is unnecessary for AI analysis.
Task-oriented communications is an approach that transmits only the information required for an AI task. Existing approaches, however, have required separate training of encoders that extract and compress information for each task, such as image classification or object recognition, creating a burden when handling multiple tasks or adding new ones.
The research team’s TACToC (Task-Adaptive Communication for Task-oriented Communications) supports multiple AI tasks with a single encoder. Instead of retraining the entire encoder for each task, TACToC changes a small prompt containing information about the task to selectively identify only the visual features needed for that task. For example, when an autonomous vehicle uses the same camera footage to identify surrounding objects or estimate their distances, the system selects and transmits the information most important for each task.
At the core of the system is a “task-describable encoder” that combines two algorithms: relevant feature estimation and prompt-based task embedding. The system predicts how much each visual feature contributes to the AI task and selects only the information that is needed, while task information is represented using a small number of trainable tokens. When a new task is added, only the corresponding tokens need to be trained. The selected information is compressed and sent to the server, which reconstructs the features required for analysis rather than the original image, allowing existing AI models to be used without modification.
In experiments using public datasets across five tasks—semantic segmentation, pose estimation, image captioning, image classification, and depth estimation—TACToC reduced communication costs by 1.5× to 14.5× compared with existing methods while keeping the loss in accuracy within three percentage points. Even when multiple tasks were processed simultaneously, the amount of transmitted data changed very little as the number of tasks increased.
Switching tasks required updating only about 1,900 parameters and took approximately 12.5 milliseconds. Compared with existing task-oriented communication methods, TACToC achieved approximately three times higher training throughput while using about one-fifth as much memory. On NVIDIA’s Jetson AGX Orin edge AI device, it processed approximately 21 frames per second.
In experiments reflecting real-world mobile network conditions, TACToC maintained stable accuracy while reducing bandwidth use by approximately 16.7% compared with H.265 video compression and approximately 41.1% compared with an existing task-oriented communication method. Because it can be deployed without modifying existing communication network architectures, the technology is expected to reduce communication overhead for autonomous vehicles, robots, drones, and intelligent CCTV systems, while also supporting applications in next-generation 6G semantic communications and Physical AI services.
“mzCache” Speeds Up AI Responses by Recovering Memory While Computing
The team also developed mzCache to address memory bottlenecks in on-device AI, where AI models run directly on smartphones. On-device AI offers advantages for privacy because data do not need to be sent to external servers, but for AI to remember previous conversations, the model and conversation history must be retained in memory. Even a compact AI model can use approximately 5–7 GB of memory for the model itself and conversation processing.
The problem arises when a user switches from AI to another app, such as a camera, game, or video application. To free up space, the operating system (OS) reclaims memory that had been used by the AI. When the user returns to the AI, responses are delayed while the data are restored. Measurements conducted by the research team on actual smartphones showed that the time to first response increased from 0.8 seconds to approximately 16 seconds. When memory pressure becomes severe, the AI may be forcibly terminated, causing the previous conversation context to be lost as well.
The researchers focused on the fact that this problem stems not only from the computational speed of the AI model itself but also from the operating system’s memory management. In practice, existing memory compression techniques reduced AI conversation history by only 0.2–9.3%.
mzCache divides AI memory into small units, allowing the system to reclaim only as much memory as other apps require. When the user returns, it performs memory recovery and AI computation simultaneously. While the graphics processing unit (GPU) performs computations using data that are already available, the central processing unit (CPU) restores the remaining data.
Taking advantage of the fact that AI models perform computations sequentially from their earlier layers, the researchers designed the system to evict memory beginning with the later layers and restore it beginning with the earlier layers. They also combined two recovery paths—one that retains compressed conversation history in memory and another that uses storage—allowing recovery to adapt to the amount of memory that has been reclaimed.
Experiments on commercially available smartphones, including the Galaxy S25+ and OnePlus 12, showed that mzCache reduced the AI’s time to first response by 2.1× to 5.5× compared with conventional methods that move data to storage and later restore it. When AI memory was completely reclaimed, mzCache delivered responses up to 25.9× faster than the default Android approach, while also reducing the energy consumed for recovery and computation.
The benefits were also confirmed in an experiment in which users alternated among 10 apps—including social media, video, games, and camera apps—for approximately 45 minutes. Under the conventional operating-system approach, the AI was forcibly terminated each time. With mzCache, however, the AI remained active throughout 10 repeated trials and preserved the context of previous conversations.
mzCache was implemented in approximately 6,000 lines of code and can run as an Android application without modifying the kernel, the core component of the operating system. It can be used to reduce response delays for AI assistants operating alongside multiple apps and is expected to expand in the future to devices such as smart glasses, laptops, and workstations that use unified memory.
Regarding the TACToC research, Professor Lee said, “This research presents a highly efficient machine-to-machine communication technique that selectively transmits only the information needed for various AI tasks, establishing a technological foundation for realizing network-based Physical AI.”
Regarding the mzCache research, he said, “This research presents a high-performance approach to operating and managing AI models that maintains both the responsiveness and conversational context of on-device AI even under limited memory.” He added, “We expect it to develop into a foundation for delivering stable AI services across a wide range of devices, including smartphones and smart glasses, and to make a significant contribution to establishing Korea as a leader in Physical AI.”
The TACToC research was conducted by Huijeong Choe, the first author, along with Minjae Lee, Taehun Kim, and Dr. Wooseung Nam of KAIST. The mzCache research involved Hongseung Yu, an M.S. student and the first author, along with Minsung Kim, Jongseok Park, and others. Professor Kyunghan Lee served as the corresponding author of both papers. The research team plans to expand the technologies to environments in which multiple AI applications run simultaneously and to a broader range of mobile processors, including neural processing units (NPUs).
Both studies were supported by the National Research Foundation of Korea and the Institute of Information & Communications Technology Planning & Evaluation, funded by the Ministry of Science and ICT. The Institute of Engineering Research at Seoul National University and the Institute of New Media and Communications also supported the research.
□ Introduction to the SNU College of Engineering
Seoul National University (SNU) founded in 1946 is the first national university in South Korea. The College of Engineering at SNU has worked tirelessly to achieve its goal of ‘fostering leaders for global industry and society.’ In 12 departments, 323 internationally recognized full-time professors lead the development of cutting-edge technology in South Korea and serving as a driving force for international development.
Caption
If AI memory is reclaimed while other apps are in use, restoring that memory can substantially increase the time to first response.
Credit
© 2026 Copyright held by the owner/author(s). Hongseung Yu et al., originally published in the Proceedings of ACM MobiCom 2026. Licensed under CC BY 4.0. https://doi.org/10.1145/3795866.3844495
No comments:
Post a Comment