We investigate a family of simple long-context attacks on large language models: prompting with hundreds of demonstrations of undesirable behavior. This attack is newly feasible with the larger context windows recently deployed by language model providers like Google DeepMind, OpenAI and Anthropic. We find that in diverse, realistic circumstances, the effectiveness of this attack follows a power law, up to hundreds of shots. We demonstrate the success of this attack on the most widely used state-of-the-art closed-weight models, and across various tasks. Our results suggest very long contexts present a rich new attack surface for LLMs.

The K L1-norm Principal Components (L1-PCs) of a data matrix X ∈ ℝD × Ncan be found optimally with cost O(2NK), in the general case, and O(Nrank(X)K - K + 1), when rankX is a constant with respect to N [1],[2]. Certainly, in real-world applications where N is large, even the latter polynomial cost is prohibitive. In this work, we present L1-BF: a novel, near-optimal algorithm that calculates the K L1-PCs of X with cost O (NDmin{N, D} + N2(K4+ DK2) + DNK3), comparable to that of standard (L2-norm) Principal-Component Analysis. Our numerical studies illustrate that the proposed algorithm attains optimality with very high frequency while, at the same time, it outperforms on the L1-PCA metric any counterpart of comparable computational cost. The outlier-resistance of the L1-PCs calculated by L1-BF is documented with experiments on dimensionality reduction and genomic data classification for disease diagnosis.

Beamforming and equalization are two key, but power hungry functions, essential in a mm-wave receiver. This paper proposes the incorporation of null adaptation and beam steering into the beamformer in order to reduce the length of the time-domain equalizer. The adaptive beamformer spatially filters the received signal field by steering main beam towards the angle of arrival of desired signal and steering nulls towards the angle of arrival of strong multipath components, thereby reducing their adverse effects prior to the DFE. This, in turn, significantly reduces the complexity and hence the power consumption of the DFE. The development of this architecture is closely guided by hardware considerations; in particular, the spatial processing hardware is implemented in the analog baseband domain, while the feedback equalizer is implemented using mixed analog-digital hardware. Adaptation circuitry is implemented using mostly digital hardware. Using this technique, higher data rates can be supported with a given number of elements. Conversely, the number of antennas and/or equalizer taps may be decreased while meeting a given data rate target. Extensive simulations using real mm-wave channel models are presented to support these hypotheses.

We address the problem of recovering an unknown image of interest, when only few, severely corrupted copies are available. We employ, for the first time in the literature, corruption-resistant L1-Principal-Components (L1-PCs) of the image data-set at hand. Specifically, the calculated L1-PCs are used for reliability-based patch-by-patch fusion of the corrupted image copies into a single high-quality representation of the original image (L1-fusion). Our experimental studies illustrate that the proposed method offers remarkable recovery results for several common corruption types, even under high corruption rate, small number of copies, and varying corruption type among copies. An additional theoretical contribution of this work is that the L1-PC of a data matrix of non-negative entries (e.g., image data) is for the first time shown to be optimally calculable with complexity linear to the matrix dimensions - as of now, the fastest-known optimal algorithm is of polynomial complexity. In the light of this result, L1-fusion is carried out with linear cost comparable to that of the simple copy-averaging alternative. The linear-low cost of L1-fusion allows for the recovered image to be, optionally, further refined by means of sophisticated single-image restoration techniques.

Rising fuel prices and environmental harmful emissions led to plethora of research for fuel and emission reduction. Using the upcoming V2V/V2I technology and the Connected Vehicle (CV) model, we address the problem of eco- driving, which examines how the latest technology can be harnessed to improve fuel-efficiency and reduce harmful emissions. In this paper, for the first time in the literature we propose microscopic eco- speed control algorithms unlike most of the existing studies in this area where the focus is on control of infrastructure side (i.e., signal timing plans). We use the concept of platoon of instrumented vehicles (IV) that are connected in a journey covering multiple intersections in a multiple vehicle setting. We propose greedy algorithms based on driving behavior and the microscopic fuel and emission models that advises a speed trajectory to minimize fuel consumption and greenhouse emissions. Detailed simulations are carried out demonstrating efficacy of the proposed algorithms for fuel and emission reductions in comparison to the naive driving behavior.

Recently, Markopoulos et al. [1], [2] presented an optimal algorithm that computes the L1maximum-projection principal component of any set of N real-valued data vectors of dimension D with complexity polynomial in N, O(ND). Still, moderate to high values of the data dimension D and/or data record size N may render the optimal algorithm unsuitable for practical implementation due to its exponential in D complexity. In this paper, we present for the first time in the literature a fast greedy single-bit-flipping conditionally optimal iterative algorithm for the computation of the L1principal component with complexity O(N3). Detailed numerical studies are carried out demonstrating the effectiveness of the developed algorithm with applications to the general field of data dimensionality reduction and direction-of-arrival estimation.

We consider arbitrary Hybrid-Automatic-Repeat-Request (H-ARQ) wireless links over quasi-static Rayleigh fading channels. In this paper, we translate the repeat-request advantage of the intended receiver over potential eavesdroppers to link security. In particular, with statistical-only knowledge of the channel and noise, we find for the first time in the literature the optimal power allocation sequence over the H-ARQ rounds that maximizes the outage probability of eavesdroppers for any given target outage probability of the trusted receiver. Simulation studies demonstrate orders of magnitude difference in outage probability between eavesdroppers and intended receiver.

We develop a novel waveform design approach to minimize the likelihood that a message transmitted wirelessly between trusted single-antenna nodes is intercepted by an eavesdropper. In particular, first, with knowledge of the eavesdropper's channel state information (CSI) we find the optimal waveform and transmit energy that minimize the signal-to-interference-plus-noise ratio (SINR) at the output of the eavesdropper's maximum-SINR linear filter, while at the same time provide the intended receiver with a required pre-specified SINR at the output of its own max-SINR filter. Next, if prior knowledge of the eavesdropper's CSI is unavailable, we design a waveform that maximizes the amount of energy available for generating disturbance to eavesdroppers, termed artificial noise (AN), while the SINR of the intended receiver is maintained at the pre-specified level. Simulation studies demonstrate our analytical developments and illustrate the benefits of the designed waveforms on securing single-input single-output (SISO) transmissions.

Pseudo-noise (PN) masking is regarded as an effective means to combat data eavesdropping (for example in military-grade communications or positioning systems). At the same time, PN-masked data transmissions are considered vulnerable to interference/jamming due to lack of practical interference suppression solutions. In this work, (i) we derive an efficient minimum-mean-square-error (MMSE) optimal linear receiver of PN-masked data and (ii) develop an auxiliary-vector (AV) MMSE adaptive filter estimator with state-of-the-art small-sample-support estimation performance. Simulation studies included in this paper illustrate the effectiveness of the theoretical developments.

Wireless physical-layer security is an emerging field of research aiming at preventing eavesdropping in an open wireless medium. In this paper, we propose a novel waveform design approach to minimize the likelihood that a message transmitted between trusted single-antenna nodes is intercepted by an eavesdropper. In particular, with knowledge first of the eavesdropper's channel state information (CSI), we find the optimum waveform and transmit energy that minimize the signal-to-interference-plus-noise ratio (SINR) at the output of the eavesdropper's maximum-SINR linear filter, while at the same time provide the intended receiver with a required pre-specified SINR at the output of its own max-SINR filter. Next, if prior knowledge of the eavesdropper's CSI is unavailable, we design a waveform that maximizes the amount of energy available for generating disturbance to eavesdroppers, termed artificial noise (AN), while the SINR of the intended receiver is maintained at the pre-specified level. The extensions of the secure waveform design problem to multiple intended receivers are also investigated and semidefinite relaxation (SDR) -an approximation technique based on convex optimization- is utilized to solve the arising NP-hard design problems. Extensive simulation studies confirm our analytical performance predictions and illustrate the benefits of the designed waveforms on securing single-input single-output (SISO) transmissions and multicasting.

We develop a new 4 × 4 Hadamard-precoded quasi-orthogonal space-time block code (QO-STBC) that enables highly effective near-maximum-likelihood (near-ML) reliability-based prioritized symbol detection using linear filters. Approximate block-error-rate minimization is being used to optimize the code rotation angle. Detailed computational complexity evaluation of the decoder in terms of real multiplications and additions shows significant complexity reduction for symbol alphabet sizes of interest. Numerical and simulation studies demonstrate negligible bit-error-rate degradation compared to the state-of-the-art in bit-error-rate by ML decoded 4 × 4 codewords.

In the context of secure wireless communications, pseudo-noise (PN) masking of transferred data has proven to be an effective technique against eavesdropping (notable examples are military-grade communication and global-positioning systems). At the same time, PN-masked transmissions are thought to be vulnerable to interference/jamming due to lack of a minimum-mean-square-error (MMSE) disturbance suppressing solution. In this paper, for the first time we establish the MMSE operation for masked data in the form of a time (mask) varying linear filter, suggest an implementation that avoids repeated input autocorrelation matrix inversion, and develop an auxiliary-vector (AV) MMSE filter estimator with state-of-the-art short-data-record estimation performance. Simulation examples included herein illustrate the theoretical developments.

We present a new 4×4 Hadamard-precoded Quasi-Orthogonal Space-Time Block Code (QO-STBC) that enables reliability-based prioritized symbol detection using linear filters. Approximate block-error-rate analysis is carried out and used to optimize the code rotation angles. For benchmarking purposes, the maximum-likelihood (ML) code detector is also derived with complexity of the order of joint two-real-symbol decoding. Numerical and simulation studies compare linear filter decoding against ML decoding.

In this paper, we present the lowest-computational-complexity maximum-likelihood (ML) decoder known to-date for 4×4 full-diversity Quasi-Orthogonal Space-Time Block Codes (QO-STBC) with symbols from square or rectangular quadrature amplitude modulation (QAM) constellations. The complexity savings come from a simplified quadratic ML decoding statistic that is being presented and the utilization of the signal points of the QAM constellation. Comparative computational complexity analysis is carried out and a simulation study demonstrates the theoretical equivalence of the proposed and original ML implementation.