We demonstrate that geometrically distinctive keys during LLM inference tend to have high attention scores. Based on the phenomenon we propose KeyDiff, a training-free KV cache eviction method based solely on key similarity. Unlike other KV cache eviction methods, KeyDiff can process arbitrarily long prompts within strict resource constraints and efficiently generate responses. We provide a theoretical basis for KeyDiff by relating key diversity with attention scores. These results imply KeyDiff can efficiently identify the most important tokens to retain. Notably KeyDiff does not rely on attention scores, allowing the use of optimized attention mechanisms like FlashAttention. Under a strict memory allowance, we demonstrate the effectiveness of KeyDiff for the Llama and Qwen model families by observing a performance gap of less than 0.04\% with 8K cache budget (~23\% KV cache reduction) from the non-evicting baseline on LongBench for Llama 3.1-8B and Llama 3.2-3B. We also observe near baseline performance for Deepseek-R1-Distill-Llama-8B on the Math500 reasoning benchmark and decrease end-to-end inference latency by up to 30\% compared to the other token-eviction methods.

Ultrasound scanning is an imaging technique that aids medical professionals in diagnostics and interventional procedures. However, a trained human-in-the-loop (HITL) with a radiologist is required to perform the scanning procedure. We seek to create a novel ultrasound system that can provide imaging in the absence of a trained radiologist, say for patients in the field who suffered injuries after a natural disaster. One challenge of automating ultrasound scanning involves finding the optimal area to scan and then performing the actual scan. This task requires simultaneously maintaining contact with the surface while moving along it to capture high quality images. In this work, we present an automated Robotic Ultrasound System (RUS) to tackle these challenges. Our approach introduces a Bayesian Optimization framework to guide the probe to multiple points on the unknown surface. Our proposed framework collects the ultrasound images as well as the pose information at every probed point to estimate regions with high vessel density (information map) and the surface contour. Based on the information map and the surface contour, an area of interest is selected for scanning. Furthermore, to scan the proposed region, a novel 6-axis hybrid force-position controller is presented to ensure acoustic coupling. Lastly, we provide experimental results on two different phantom models to corroborate our approach.

A robotic swarm can perform various tasks. However, a human is required to task the swarm. Human control over the swarm can be enabled through a set of influential agents which can be either leaders or predators. In the presence of multiple tasks, the swarm may need to split into sub-swarms to accomplish the task and re-group as a swarm to execute larger tasks. The response of the swarm in the presence of influential agents depends on the swarm dynamics. A precise measure of influence using leaders or predators or a combination of leaders and predators to achieve the mission is not adequately studied. In this paper, we analyze the effect of using only leaders, only predators and a combination of leaders and predators on three swarm models namely, shepherding model, Couzin's model and a physicomimetic models while they perform foraging tasks and carry out Monte-Carlo simulations to evaluate the performance of the influential agents on different swarms. We also propose a novel way to split a swarm into smaller sub-swarms using influential agents. Our results show that the predator based swarm splitting and steering to a task based on shepherding model performs far better than any other combination of leaders and predators. This result is consistent even when the number of agents is increased to 500.