For decades, smart environments have promised kitchens that know when appliances are running, workshops that track which tools are in use, and buildings that sense how people move through them. Yet most buildings remain largely ''dumb.'' The obstacles are familiar: sensors require batteries and maintenance, the wireless spectrum is crowded, and continuously operating cameras and radios consume more power than perpetual deployment can afford. All of these obstacles share a common root: conventional sensing systems sample the world continuously, burning power and bandwidth even when nothing is happening. Biology solved this problem differently. The retina in the human eye doesn't refresh at a fixed rate. It responds asynchronously to change, staying quiet until something moves and then signaling only where and when it did. What if our sensing infrastructure worked the same way?
Eye tracking is becoming a central input modality for extended-reality and smart-glasses systems. In virtual reality, gaze can support foveated rendering, avatar eye contact, and hands-free selection. On AR or everyday smart glasses, gaze has an additional role: it tells what the wearer is paying attention to in the surrounding world. When a user asks, ''What is this?'', the answer depends on whether the user is looking at a sign, a warning light, a product label, or a person in the scene. This context is especially important for on-device AI agents. A front-facing camera can capture the scene, but it does not by itself reveal the user's attention. Gaze can disambiguate speech, focus perception, and enable low-effort interaction without asking the wearer to point or touch a controller. The challenge is that smart glasses have a much tighter hardware budget than headsets. Sensors must be small, low power, robust to outdoor conditions, and unobtrusive enough to fit into eyewear.
Long-context language models are attractive because they can read whole documents, code repositories, conversations, and scientific sequences. However, a long inference window is not the same as a model that was trained to use millions of tokens reliably. Direct long-context training is blocked by temporary memory spikes from activations, attention buffers, communication workspaces, and the final loss computation, even when model parameters and optimizer states are already sharded. Our proposed method, Fully Pipelined Distributed Transformer (FPDT), treats the long sequence problem as a memory-hierarchy scheduling problem. It chunks each sequence before distributed attention, stages inactive query/ key/value tensors in host memory and uses double buffers to overlap transfers with attention computation. FPDT democratizes long sequence training, making it accessible on mobile and commodity hardware, such as high-end local machines. In our evaluation, FPDT trains contexts up to 16x longer than prior sequence-parallel baselines on the same hardware. It reaches a 2M-token context for an 8B model on four A100-80GB GPUs, and a 4M-token context for 8B Llama 3 on eight A100-80GB GPUs while sustaining 55.7% Model FLOPs Utilization (MFU) [1].
Blood pressure is central to cardiovascular assessment, yet routine measurements are uncomfortable to repeat and usually provide only sparse systolic and diastolic readings. WaveBP explores whether a commercial mmWave radar can recover the arterial pressure waveform beat by beat without attaching a sensor to the body.
Wireless signals are invisible, and their propagation through the physical environment is inherently complex. Walls, furniture, people, and other objects block, reflect, diffract, and scatter radio waves, creating strong spatial variations in signal quality. Modern wireless systems further increase this complexity through wider bandwidth and directional beamforming. Since signal quality directly affects link reliability, throughput, and overall network performance, understanding radio propagation is essential for network deployment and debugging. Yet engineers typically observe only measurements such as signal strength, throughput, or packet loss at sampled locations, leaving them to infer the underlying spatial propagation process from sparse measurements and repeated experiments.
