Two new systems out of Seoul National University attack the twin bottlenecks of practical AI: AI data traffic, the bandwidth wasted moving information machines do not need, and response delays, the minutes a phone takes to wake an on-device assistant after the operating system has reclaimed its memory.
Both papers were presented at ACM MobiCom 2026, the premier mobile systems conference, and both take the same quiet position: stop treating every AI workload as if it were a human watching video or a human waiting for an app. TACToC cuts AI data traffic by up to fourteen and a half times by transmitting only the visual facts a task actually requires. mzCache revives on-device AI up to twenty-five point nine times faster by recovering memory and computation at the same time.
Neither system replaces an AI model. Neither needs new hardware. Together they sketch a future in which cameras, robots, phones and edge servers stop shouting everything they know at each other, and stop freezing when memory runs short. This article breaks down how each system works, what the published numbers say, and why the pattern matters well beyond the lab.
1. Cutting AI Data Traffic And Response Delays
2. TACToC: One Encoder For Many Tasks
3. The Numbers: 14.5x Less Traffic, Three Points Lost
4. Where Task-Adaptive Communication Pays Off
5. On-Device AI’s Memory Problem
6. mzCache: Recover While You Compute
7. Real Phones, Real Multitasking
8. What This Means For Physical AI
References
Cutting AI Data Traffic And Response Delays
Two Papers, One Shared Enemy
Both systems come from the same lab orbit. Professor Kyunghan Lee of Seoul National University’s Department of Electrical and Computer Engineering is the corresponding author on the two papers that ACM MobiCom 2026 accepted from the group, presented in Austin, Texas between October 26 and 30. MobiCom is the flagship venue for mobile systems research, the place where papers must survive implementation on real devices, and both of these proposals do exactly that.
The enemy they share is waste. Modern AI keeps getting smarter, but the plumbing around it still behaves like plumbing from another decade: cameras flood networks with pixels nobody analyzes, and phones quietly evict the memory their assistants need. Cut the first waste and you cut AI data traffic. Cut the second and you cut the response delays that make mobile AI feel broken.
Why The Two Problems Travel Together
They are two halves of one story about moving and holding. When a phone, a drone or an autonomous vehicle offloads inference to a server, the bottleneck is the network: frames must be captured, compressed and transmitted, and conventional compression optimizes for the human eye, not the machine analyzing the stream. That mismatch is where AI data traffic balloons.
When instead the model runs on the device itself, the bottleneck moves to memory. The model and its conversation history must live in RAM while other apps compete for the same chips. Under pressure the operating system reclaims that memory, and the user pays in seconds of latency, or in a lost conversation. TACToC fixes the moving side; mzCache fixes the holding side.
There is a third connection the authors make explicit. Every on-device failure that pushes work to the cloud feeds AI data traffic back into the network, so fixing memory churn on phones and trimming AI data traffic at the edge are two fronts of one campaign to make machine intelligence affordable to run every day.
What makes the pair compelling is arithmetic rather than aspiration. MobiCom reviewers saw prototypes measured against stock pipelines, on datasets the community can re-check, and the reductions in AI data traffic and wake-up delay replicated across tasks, phones and network conditions. Numbers that survive contact with real hardware are rare in AI publishing; here are two in one session.
TACToC: One Encoder For Many Tasks
What Task-Oriented Communication Already Does
TACToC — short for Task-Adaptive Communication for task-oriented communications — belongs to a research line called task-oriented communication. The premise is simple: when a machine looks at a video feed, it does not need the picture. It needs conclusions. Classic systems push that idea only so far, because each task, say classification or object recognition, demands its own encoder and decoder pairing. Add a task, rebuild the pipeline, and watch AI data traffic creep back up as every stream keeps its own bespoke channel.
The Task-Describable Encoder
TACToC’s central trick is a single encoder that serves many tasks at once. The encoder changes a small prompt carrying the task description, and that prompt steers which visual features get packed into the transmission. One autonomous vehicle can use the same camera footage to identify objects at one moment and estimate their distance the next, without re-encoding the world from scratch each time.
Under the hood the team built what they call a task-describable encoder, composed of relevant feature estimation and prompt-based task embedding. The system predicts how much each visual feature contributes to the active task and selects only the useful information. Task descriptions themselves are small trainable tokens, so adding a new task updates roughly 1,900 parameters and switches tasks in about 12.5 milliseconds.
That economics is transformative for fleets. When adding recognition of a new road hazard costs almost no training time and negligible AI data traffic, an operator can afford dozens of tasks on one camera. TACToC turns AI data traffic from a fixed tax per video stream into a per-question charge, closer to paying for answers than for raw reality.
The Numbers: 14.5x Less Traffic, Three Points Lost
Five Tasks, One Encoder
The team evaluated TACToC on public datasets across five quite different tasks: semantic segmentation, pose estimation, image captioning, image classification and depth estimation. Across that spread, communication costs fell between 1.5 and 14.5 times compared with existing methods, while accuracy loss stayed within three percentage points. Crucially, transmitted data barely grew as more tasks ran simultaneously — the property that separates a task-adaptive design from a pile of per-task pipelines.
Training economics improved too. TACToC trained with roughly three times the throughput and about one fifth of the memory that existing task-oriented methods require, and it reached around 21 frames per second on an NVIDIA Jetson AGX Orin, an edge module small enough to sit inside a vehicle.
Every one of those reductions is a dent in AI data traffic a fleet pays for monthly: less spectrum consumed, fewer congested radio chains, smaller invoices. The five-task spread suggests the AI data traffic savings hold even when several analyses share one stream, which is exactly how robots and cameras behave in production. The team also expects the gains to widen with time, because the encoder learns which features matter per task and stops paying for the rest.
Measured Against The Network Itself
In real mobile network conditions the system cut bandwidth by about 16.7 percent against H.265 video while holding accuracy, and by roughly 41.1 percent against a prior task-oriented method. Those margins matter most at the cell edge, where a congested radio turns every extra kilobyte of AI data traffic into visible stutter for the machine on the other end of the link.
Table 1 gathers the headline figures from the MobiCom paper in one glance.
| Metric | Result |
|---|---|
| Communication cost reduction | 1.5x to 14.5x versus existing methods |
| Accuracy loss | Within 3 percentage points |
| New-task parameters | About 1,900 trainable token parameters |
| Task switch time | About 12.5 milliseconds |
| Edge speed | About 21 fps on Jetson AGX Orin |
| Bandwidth versus H.265 | Down about 16.7 percent |
| Bandwidth versus prior task-oriented method | Down about 41.1 percent |
Where Task-Adaptive Communication Pays Off
Because The Server Sees Features, Not Pictures
One architectural choice makes TACToC unusually deployable. The server reconstructs the features themselves for analysis rather than rebuilding the original image, which means existing AI models consume the received information without modification. No fine-tuning parade, no new model zoo — trimming AI data traffic becomes a drop-in upgrade to the transport layer.
The team also stresses that nothing about current network architecture must change. The savings apply under real mobile conditions with fluctuating bandwidth, precisely the environment where autonomous vehicles, delivery robots, drones and intelligent CCTV choke first when connectivity degrades.
Fed Into The 6G Semantic Story
Professor Lee frames the result as a “highly efficient machine-to-machine communication technique that selectively transmits only the information needed for various AI tasks,” one that establishes “a technological foundation for realizing network-based physical AI.” That phrasing matters: 6G standardization has been circling the idea of semantic communication for years, networks that carry meaning rather than bits. TACToC is the kind of working, measured system those white papers need, and the authors explicitly position it as groundwork for 6G semantic communication and for physical AI at scale.
For buyers the pitch is plain: lower AI data traffic means cheaper plans, lighter radios and longer battery life on every connected machine. When a camera robot sends one-tenth the AI data traffic for the same analysis, whole product categories — always-on fleet vision, city-scale CCTV analytics — move from impossible to merely expensive.
Regulators notice too. A system that ships fewer identifiable pixels over public networks carries less privacy exposure by construction, another quiet dividend of reduced AI data traffic that operators of public cameras will not ignore when they shortlist next-generation video links.
On-Device AI's Memory Problem
Privacy Costs Gigabytes
Running AI on the device looks like the obvious answer to a dozen complaints at once. Nothing leaves the phone, so privacy holds; there is no round trip, so latency should vanish. But a model and its conversation history must stay resident in RAM, and even a compact on-device AI occupies roughly five to seven gigabytes once the model and its conversation processing are counted. On a phone where every app fights for the same memory, that is a boulder in a bathtub.
The 16-Second Resurrection
The consequences are measurable and ugly. When you switch apps, the operating system reclaims the memory the AI was holding, because your assistant obviously has no more claim on RAM than a game does. Return to the assistant and it must rebuild what it lost before answering a single word.
The SNU team measured time-to-first-response on real phones growing from about 0.8 seconds to roughly 16 seconds as memory pressure climbed — a twentyfold penalty hiding behind the same friendly icon. In severe cases the system force-kills the AI outright and the conversation context disappears with it. Existing memory-compression tricks offered little rescue: they shrank conversation history by only 0.2 to 9.3 percent, nowhere near enough to defuse the collision.
And the failure mode has an AI data traffic shadow. A killed assistant that re-warms from the cloud, or a frustrated user whose query gets retried through a hosted service, pushes traffic back onto the network the TACToC crowd is trying to quiet. Wasted phone memory, in other words, eventually becomes wasted AI data traffic too.
That double bill explains the urgency behind both MobiCom papers. Hardware budgets are not loosening anytime soon, so the next generation of mobile assistants will either learn to hold on tighter and speak softer — less memory churn and less AI data traffic — or keep losing arguments to whatever game just opened in the background.
mzCache: Recover While You Compute
Small Units, Selective Reclaim
mzCache reorganizes how an on-device AI stores its own memory. Instead of one monolithic block the operating system can only seize wholesale, the AI’s state is divided into small units. When other apps need room, the system reclaims only as much as they actually need, leaving the rest of the assistant warm. It is the holding-side mirror of TACToC’s idea: keep or send only what matters.
Two Engines, One Clock
The recovery design exploits a structural fact about neural networks: they compute layer after layer in sequence. mzCache evicts memory from later layers first and restores from earlier layers first, so by the time computation walks back through the network, the pieces it needs have already been put home.
Then the overlap trick. On return, recovery and computation run simultaneously — the GPU starts computing on the memory that is already present while the CPU restores the remainder in parallel. Two dual recovery paths, one holding compressed history in memory and one rebuilding from storage, adapt to how much memory was actually taken, so the system never pays for a restoration it does not need.
The design philosophy rhymes with TACToC’s across the whole stack. A task-adaptive encoder prunes AI data traffic on the wire; a memory-zoned runtime prunes the churn inside RAM. Both treat selectivity as the cheapest performance there is, and both keep AI data traffic and AI memory footprints shrinking as more tasks and more apps pile on.
One subtlety deserves applause: nothing in mzCache asks the cloud for help. The revival happens locally, so the user waiting for an answer generates no extra AI data traffic while the system heals itself — patience bought with bytes that were never spent.
Real Phones, Real Multitasking
Galaxy And OnePlus, Not Emulators
Evaluation happened on a Galaxy S25+ and a OnePlus 12, and the margins were dramatic. mzCache cut time-to-first-response by 2.1 to 5.5 times compared with storage-based recovery, and when memory had been fully reclaimed it finished up to 25.9 times faster than default Android. The energy of recovery plus computation dropped as well, which matters as much as speed on a phone you need to survive the evening.
The Ten-App Gauntlet
The most telling experiment alternated the assistant with ten applications — social media, video, games and the camera — for around 45 minutes. On the conventional Android path, the system force-killed the AI every single time; users lost their conversational context repeatedly across the run. With mzCache the assistant stayed alive through all ten trials, context intact, ready to answer the moment it was summoned.
None of this required heroic cooperation from the phone vendor either. mzCache is about 6,000 lines of code that runs as an Android application without modifying the kernel, a systems-hack elegance that makes it shippable far faster than a platform rewrite would be — and every kept-alive assistant is one fewer cloud retry adding AI data traffic when the user returns.
Table 2 contrasts the two behaviors side by side.
| Behavior | Default Android | With mzCache |
|---|---|---|
| Wake-up after memory reclaim | Up to about 16 seconds | 2.1x to 5.5x faster to first response |
| Fully reclaimed memory | Full rebuild from cold | Up to 25.9x faster |
| Ten-app, 45-minute session | AI force-killed every time | AI alive all ten trials |
| Conversation context | Lost on force-kill | Preserved |
| Kernel modification | Not applicable | None needed |
What This Means For Physical AI
The Same Scarcity Nvidia Keeps Hitting
These papers land exactly where the industry’s biggest bottlenecks sit. Memory is the choke point the entire sector keeps running into; Nvidia’s own engineers describe the memory wall as the defining constraint of AI infrastructure, and squeezing AI data traffic on the edge is the mirror image of that fight inside the datacenter. Meanwhile the phone and PC markets feel the same squeeze from the silicon side, where the AI chip supply crisis keeps making every gigabyte of RAM more contested.
Efficiency Is Becoming The Whole Game
Efficiency pressure is no longer niche. Every wasted token, pixel and byte now shows up in energy bills, in measurements of how much more energy models burn on some inputs than others, and in founders’ invoices. Cutting AI data traffic is therefore being recognized as an efficiency lever equal in stature to model distillation: you cannot always shrink the model, but you can always stop moving the model’s homework around the planet.
Professor Lee casts mzCache in exactly this national-technology light, calling it a high-performance way to operate AI that “maintains both the responsiveness and conversational context of on-device AI even under limited memory,” and saying work like this will “contribute to establishing Korea as a leader in physical AI.”
The team’s next steps are practical: multi-AI-app environments, broader mobile processors including NPUs, and mzCache-style memory management spreading from phones to smart glasses, laptops and workstations with unified memory. Watch for the first vendor who advertises AI data traffic per query as a headline spec — that product manager will be reading MobiCom proceedings this year.
If TACToC taught the network to listen for the point, and mzCache taught the phone to remember on a budget, the lesson for 2027 is blunt: cutting AI data traffic and AI response delays stops being a research demo and becomes a product requirement for every company shipping physical AI hardware.
References
- Seoul National University, “New communication and memory tools reduce AI data traffic and response delays,” TechXplore, November 5, 2026. techxplore.com
- Huijeong Choe, Minjae Lee, Taehun Kim, Wooseung Nam, Kyunghan Lee, “TACToC: Task-Adaptive Communication for Task-Oriented Communications,” ACM MobiCom 2026. doi.org/10.1145/3795866.3796689
- Hongseung Yu, Minsung Kim, Jongseok Park, Kyunghan Lee, mzCache memory-management paper for on-device AI, ACM MobiCom 2026. doi.org/10.1145/3795866.3844495
- ACM MobiCom 2026 conference site. sigmobile.org/mobicom/2026
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
