KV cache compression promises increased throughput and efficiency with negligible loss in performance.While the gains in throughput are indisputable and recent literature has indeed shown minimal degradation on particular benchmarks, in general the consequences of compression in realistic scenarios such as multiinstruction prompting have been insufficiently studied.In this paper, we identify several pitfalls that practitioners should be aware of when deploying KV cache compressed LLMs.We evaluate five KV cache compression methods (StreamingLLM, SnapKV, TOVA, H2O, and K-Norm) on Llama3 8B and Qwen2.5 14B under multi-instruction prompting with IFEval.Importantly, we show that certain instructions degrade much more rapidly with compression, effectively causing them to be completely ignored by the LLM.As a practical example, we highlight system prompt leakage as a case study, empirically demonstrating the impact of compression on leakage and general instruction-following.We identify several factors that contribute to system prompt leakage: compression method, instruction order, and KV eviction bias.We then propose simple changes to KV cache eviction policies that can reduce the impact of these factors and improve the overall performance in multi-instruction tasks.alexluchen/pitfalls-of-kv-cache-compression

Chart corpora, which comprise data visualizations and their semantic labels, are crucial for advancing visualization research. However, the labels in most existing corpora are high-level (e.g., chart types), hindering their utility for broader applications in the era of AI. In this paper, we contribute VisAnatomy, a corpus containing 942 real-world SVG charts produced by over 50 tools, encompassing 40 chart types and featuring structural and stylistic design variations. Each chart is augmented with multi-level fine-grained labels on its semantic components, including each graphical element's type, role, and position, hierarchical groupings of elements, group layouts, and visual encodings. In total, VisAnatomy provides labels for more than 383k graphical elements. We demonstrate the richness of the semantic labels by comparing VisAnatomy with existing corpora. We illustrate its usefulness through four applications: semantic role inference for SVG elements, chart semantic decomposition, chart type classification, and content navigation for accessibility. Finally, we discuss research opportunities to further improve VisAnatomy.

Hyperbolic space naturally encodes hierarchical structures such as phylogenies (binary trees), where inward-bending geodesics reflect paths through least common ancestors, and the exponential growth of neighborhoods mirrors the super-exponential scaling of topologies. This scaling challenge limits the efficiency of Euclidean-based approximate Bayesian inference methods. Motivated by the geometric connections between trees and hyperbolic space, we develop novel hyperbolic extensions of two sequential search algorithms: Combinatorial and Nested Combinatorial Sequential Monte Carlo (\textsc{Csmc} and \textsc{Ncsmc}). Our approach introduces consistent and unbiased estimators, along with variational inference methods (\textsc{H-Vcsmc} and \textsc{H-Vncsmc}), which outperform their Euclidean counterparts. Empirical results demonstrate improved speed, scalability and performance in high-dimensional Bayesian phylogenetic inference tasks.

The assumption of independence between observations (units) in a dataset is prevalent across various methodologies for learning causal graphical models. However, this assumption often finds itself in conflict with real-world data, posing challenges to accurate structure learning. We propose a decorrelation-based approach for causal graph learning on dependent binary data, where the local conditional distribution is defined by a latent utility model with dependent errors across units. We develop a pairwise maximum likelihood method to estimate the covariance matrix for the dependence among the units. Then, leveraging the estimated covariance matrix, we develop an EM-like iterative algorithm to generate and de-correlate samples of the latent utility variables, which serve as de-correlated data. Any standard causal discovery method can be applied on the de-correlated data to learn the underlying causal graph. We demonstrate that the proposed de-correlation approach significantly improves the accuracy in causal graph learning, through numerical experiments on both synthetic and real-world datasets.

WiFi communication should be possible only between devices inside the same network. However, we find that all existing WiFi devices send back acknowledgments (ACKs) to even fake packets received from unauthorized WiFi devices outside of their network. Moreover, we find that an unauthorized device can manipulate the power-saving mechanism of WiFi radios and keep them continuously awake by sending specific fake beacon frames to them. Our evaluation of over 5000 devices from 186 vendors confirms that these are widespread issues. We believe these loopholes cannot be prevented, and hence they create privacy and security concerns. Finally, to show the importance of these issues and their consequences, we implement and demonstrate two attacks where an adversary performs battery drain and WiFi sensing attacks just using a tiny WiFi module which costs less than ten dollars.

Queuing delay is a dynamic network parameter that plays an important role in defining the performance of Internet applications over an end-to-end path. However, measurement of queuing delay is challenging because it requires a large infrastructural support from the path under test. In this paper, we propose an active scheme to measure queuing delay on a router using a probe-gap model. The scheme uses a popular data-clustering algorithm to process its data samples; therefore, its measurement efficacy is not dependent on the issues related to infrastructural access, certain variations (e.g., compression) in the probe gaps, and the number of clusters in the data processing. Here, we present a detailed evaluation of the scheme against the current state-of-the-art on a single-hop path through ns-3 simulation. Our results show that the proposed scheme is robust, consistent, quick, and highly accurate under different traffic conditions.

Recent advances in IoT, machine learning and cloud computing have placed a huge strain on wireless networks. In particular, many emerging applications require streaming rich content (such as videos) in real time, while they are constrained by energy sources. A wireless network which supports high data-rate while consuming low-power would be very attractive for these applications. Unfortunately, existing wireless networks do not satisfy this requirement. For example, WiFi backscatter and Bluetooth networks have very low power consumption, but their data-rate is very limited (less than a Mbps). On the other hand, modern WiFi and mmWave networks support high throughput, but have a high power consumption (more than a watt).

Backscatter networks (such as RFID, and WiFi backscatter) are very attractive for IoT applications due to their ultra-low energy consumption. In fact, their required energy to operate is low enough that it can be harvested from the environment without having a battery. However, existing backscatter networks offer very limited data-rates (i.e. at most one Mbps). Hence, despite their energy benefit, their applications are very limited. This paper presents the design of mmTag, a backscatter network which can achieve Gbps data-rates. mmTag achieves this by developing a backscatter technology operating in the mmWave spectrum band. mmWave promises to enable high throughput wireless links by offering massive chunks of high-frequency spectrum. However, to use mmWave frequencies in backscatter networks, we need to address a fundamental challenge: beam alignment. mmWave devices require highly directional antennas with very narrow beams, and communication is possible only when the transmitter's beam is aligned with the receiver's beam. However, existing beam searching techniques require power hungry components, and most importantly require the node to transmit a signal which is not possible for a backscatter device. mmTag solves this problem by building a mmWave backscatter tag which performs beam alignment without using any active component. Finally, we implement mmTag and empirically demonstrate some results.

An interactive advertising system is proposed comprising of a flat screen display, a collocated sensing camera and an outcome feedback camera. The display presents an Avatar that has the goal of convincing people to enter a store. The Avatar has a repertoire of actions that it can perform such as offer a coupon or present a piece of merchandise. A sensing camera is used to represent the state of a possible customer. State variables include affective pose, facial expressions, gaze direction and eyeball movement. The outcome feedback camera is used to detect if and when a customer elects to enter the store. Given access to such feedback, a reinforcement learning algorithm is then used to construct a policy that indicates which actions should be taken as a function of observed customer state. The efficacy of this approach is considered via synthetic experiments as well as real world demonstrations. The paper includes a discussion regarding new descriptors based on emergent languages that can be used to augment the state space representation.

Cells transition from spread to rounded morphologies in diverse physiological contexts including mitosis and mesenchymal-to-amoeboid transitions. When these drastic shape changes occur rapidly, cell volume and surface area are approximately conserved. Consequently, the rounded cells are suddenly presented with a several-fold excess of cell surface whose area far exceeds that of a smooth sphere enclosing the cell volume. This excess is stored in a population of bleb-like protrusions (BLiPs), whose size distribution is shown by electron micrographs to be skewed. We introduce three complementary models of rounded cell morphologies with a prescribed excess surface area. A 2D Hamiltonian model provides a mechanistic description of how discrete attachment points between the cell surface and cortex together with surface bending energy can generate a morphology that satisfies a prescribed excess area and BLiP number density. A 3D random seed-and-growth model simulates efficient packing of BLiPs over a primary rounded shape, demonstrating a pathway for skewed BLiP size distributions that recapitulate 3D morphologies. Finally, a phase field model (2D and 3D) posits energy-based constitutive laws for the cell membrane, nematic F-actin cortex, interior cytosol, and external aqueous medium. The cell surface is equipped with a spontaneous curvature function, a proxy for the cell surface-cortex couple, that is a priori unknown, which the model "learns" from the thin section transmission electron micrograph image (2D) or the "seed and growth" model image (3D). Converged phase field simulations predict self-consistent amplitudes and spatial localization of pressure and stress throughout the cell for any posited stationary morphology target and cell compartment constitutive properties. The models form a general framework for future studies of cell morphological dynamics in a variety of biological contexts.