Search

Although language model (LM) agents are demonstrating growing potential in many domains, their success in cybersecurity has been limited due to simplistic design and the lack of fundamental features for this domain. We present EnIGMA, an LM agent for autonomously solving Capture The Flag (CTF) challenges. EnIGMA introduces new Agent-Computer Interfaces (ACIs) to improve the success rate on CTF challenges. We establish the novel Interactive Agent Tool concept, which enables LM agents to run interactive command-line utilities essential for these challenges. Empirical analysis of EnIGMA on over 350 CTF challenges from three different benchmarks indicates that providing a robust set of new tools with demonstration of their usage helps the LM solve complex problems and achieves state-of-the-art results on the NYU CTF and Intercode-CTF benchmarks. Finally, we discuss insights on ACI design and agent behavior on cybersecurity tasks that highlight the need to adapt real-world tools for LM agents.

We consider high-dimensional regression with a count response modeled by Poisson or negative binomial generalized linear model (GLM). We propose a penalized maximum likelihood estimator with a properly chosen complexity penalty and establish its adaptive minimaxity across models of various sparsity. To make the procedure computationally feasible for high-dimensional data we consider its LASSO and SLOPE convex surrogates. Their performance is illustrated through simulated and real-data examples.

In this paper we show how the superquadratic functions can be used as a tool for researching other types of convex functions like $\phi $-convexity, strong-convexity and uniform convexity. We show how to use inequalities satisfied by superquadratic functions and how to adapt the technique used to get them in order to obtain new results satisfied by uniformly convex functions and to $\phi $-convex functions. Also, we show examples that emphasize relations between superquadracity and some other types of convex functions, Comment: 12 pages

In recent years, notable advancements have been made in the domain of visual document understanding, with the prevailing architecture comprising a cascade of vision and language models. The text component can either be extracted explicitly with the use of external OCR models in OCR-based approaches, or alternatively, the vision model can be endowed with reading capabilities in OCR-free approaches. Typically, the queries to the model are input exclusively to the language component, necessitating the visual features to encompass the entire document. In this paper, we present VisFocus, an OCR-free method designed to better exploit the vision encoder's capacity by coupling it directly with the language prompt. To do so, we replace the down-sampling layers with layers that receive the input prompt and allow highlighting relevant parts of the document, while disregarding others. We pair the architecture enhancements with a novel pre-training task, using language masking on a snippet of the document text fed to the visual encoder in place of the prompt, to empower the model with focusing capabilities. Consequently, VisFocus learns to allocate its attention to text patches pertinent to the provided prompt. Our experiments demonstrate that this prompt-guided visual encoding approach significantly improves performance, achieving state-of-the-art results on various benchmarks., Comment: 32 pages, 18 figures

In this paper we prove results on the difference between a normalized Jensen functional and the sum of other normalized Jensen functionals for convex function., Comment: 10 pages

In this paper we improve results related to Normalized Jensen Functional for convex functions and Uniformly Convex Functions., Comment: 9 pages

The problems of uniform linear array (with uniform mutual coupling) calibration and Toeplitz covariance matrix estimation are re-examined for application in the receive arrays of modern High Frequency Over-the-Horizon Radars (HF OTHR). Specifics of these arrays, operating predominantly in the "over-sampled" (d/$\lambda$ < 0.5) regime, are exploited along with the recent findings in the Toeplitz Inverse Eigenvalue problem (ToIEP), Random Matrix Theory, and the "Expected Likelihood" methodology developed earlier to advance the problem solution. In this paper we re-evaluate the benchmark efficiency of the "blind" calibration and analyze several techniques that are based on certain properties of the Toeplitz covariance matrix and "over-sampled" uniform antenna arrays, but do not require matrix estimation. In Part II we analyze the case of the unknown symmetric and Hermitian Toeplitz matrices and introduce techniques for joint antenna calibration and Toeplitz covariance matrix estimation., Comment: We do not have permission from our sponsor to submit this paper to arXiv

Complex quantum systems and their various applications are susceptible to noise of coherent and incoherent nature. Characterization of noise and its sources is an open, key challenge in quantum technology applications, especially in terms of distinguishing between inherent incoherent noise and systematic coherent errors. In this paper, we study a scheme of repeated sequential measurements that enables the characterization of incoherent errors by reducing the effects of coherent errors. We demonstrate this approach using a coherently controlled Nitrogen Vacancy in diamond, coupled to both a natural nuclear spin bath (non-Markovian) and to experimentally controlled relaxation through an optical pumping process (nearly Markovian). Our results show mitigation of coherent errors both for Markovian and Non-Markovian incoherent noise profiles. We apply this scheme to the estimation of the dephasing time ($T_2^*$) due to incoherent noise. We observe an improved robustness against coherent errors in the estimation of dephasing time ($T_2^*$) compared to the standard (Ramsey) measurement.

We consider a deep neural network estimator based on empirical risk minimization with l_1-regularization. We derive a general bound for its excess risk in regression and classification (including multiclass), and prove that it is adaptively nearly-minimax (up to log-factors) simultaneously across the entire range of various function classes.

The cornerstone of structural biology is the unique relationship between protein sequence and the 3D structure at equilibrium. Although intrinsically disordered proteins (IDPs) do not fold into a specific 3D structure, breaking this paradigm, some IDPs exhibit large-scale organization, such as liquid-liquid phase separation. In such cases, the structural plasticity has the potential to form numerous self-assembled structures out of thermal equilibrium. Here, we report that high-temperature incubation time is a defining parameter for micro and nanoscale self-assembly of resilin-like IDPs. Interestingly, high-resolution scanning electron microscopy micrographs reveal that an extended incubation time leads to the formation of micron-size rods and ellipsoids that depend on the amino acid sequence. More surprisingly, a prolonged incubation time also induces amino acid composition-dependent formation of short-range nanoscale order, such as periodic lamellar nanostructures. We can correlate the lamellar structures to \b{eta}-sheet formation and demonstrate similarities between the observed nanoscopic structural arrangement and spider silk. We, therefore, suggest that regulating the period of high-temperature incubation, in the one-phase regime, can serve as a unique method of controlling the hierarchical self-assembly mechanism of structurally disordered proteins., Comment: 27pages, 8 figures

"Toeplitzification" or "redundancy (spatial) averaging", the well-known routine for deriving the Toeplitz covariance matrix estimate from the standard sample covariance matrix, recently regained new attention due to the important Random Matrix Theory (RMT) findings. The asymptotic consistency in the spectral norm was proven for the Kolmogorov's asymptotics when the matrix dimension N and independent identically distributed (i.i.d.) sample volume T both tended to infinity (N->inf, T->inf, T/N->c > 0). These novel RMT results encouraged us to reassess the well-known drawback of the redundancy averaging methodology, which is the generation of the negative minimal eigenvalues for covariance matrices with big eigenvalues spread, typical for most covariance matrices of interest. We demonstrate that for this type of Toeplitz covariance matrices, convergence in the spectral norm does not prevent the generation of negative eigenvalues, even for the sample volume T that significantly exceeds the covariance matrix dimension (T >> N). We demonstrate that the ad-hoc attempts to remove the negative eigenvalues by the proper diagonal loading result in solutions with the very low likelihood. We demonstrate that attempts to exploit Newton's type iterative algorithms, designed to produce a Hermitian Toeplitz matrix with the given eigenvalues lead to the very poor likelihood of the very slowly converging solution to the desired eigenvalues. Finally, we demonstrate that the proposed algorithm for restoration of a positive definite (p.d.) Hermitian Toeplitz matrix with the specified Maximum Entropy spectrum, allows for the transformation of the (unstructured) Hermitian maximum likelihood (ML) sample matrix estimate in a p.d. Toeplitz matrix with sufficiently high likelihood.

Quantifying the contribution of database facts to query answers has been studied as means of explanation. The Banzhaf value, originally developed in Game Theory, is a natural measure of fact contribution, yet its efficient computation for select-project-join-union queries is challenging. In this paper, we introduce three algorithms to compute the Banzhaf value of database facts: an exact algorithm, an anytime deterministic approximation algorithm with relative error guarantees, and an algorithm for ranking and top-$k$. They have three key building blocks: compilation of query lineage into an equivalent function that allows efficient Banzhaf value computation; dynamic programming computation of the Banzhaf values of variables in a Boolean function using the Banzhaf values for constituent functions; and a mechanism to compute efficiently lower and upper bounds on Banzhaf values for any positive DNF function. We complement the algorithms with a dichotomy for the Banzhaf-based ranking problem: given two facts, deciding whether the Banzhaf value of one is greater than of the other is tractable for hierarchical queries and intractable for non-hierarchical queries. We show experimentally that our algorithms significantly outperform exact and approximate algorithms from prior work, most times up to two orders of magnitude. Our algorithms can also cover challenging problem instances that are beyond reach for prior work.

The frozen mode regime is a unique slow-light scenario in periodic structures, where the flat-bands (zero group velocity) are associated with the formation of high-order stationary points (aka exceptional points). The formation of exceptional points is accompanied by enhancement of various optical properties such as gain, Q-factor and absorption, which are key properties for the realization of wide variety of devices such as switches, modulators and lasers. Here we present and study a new integrated optical periodic structure consisting of three waveguides coupled via micro-cavities and directional coupler. We study this design theoretically, demonstrating that a proper choice of parameters yields a third order stationary inflection point (SIP). We also show that the structure can be designed to exhibit two almost-overlapping SIPs at the center of the Brillouin Zone. We study the transmission and reflection of light propagating through realistic devices comprising a finite number of unit-cells and investigate their spectral properties in the vicinity of the stationary points. Finally, we analyze the lasing frequencies and threshold level of finite structures (as a function of the number of unit-cells) and show that it outperforms conventional lasers utilizing regular band edge lasing (such as DFB lasers)., Comment: 11 pages, 10 figures

We give a formula for the integral Chow rings of weighted blow-ups. Along the way, we also compute the integral Chow rings of weighted projective stack bundles, a formula for the Gysin homomorphism of a weighted blow-up, and a generalization of the splitting principle. In addition, in the appendix we compute the Chern class of a weighted blow-up., Comment: by: Veronica Arena, Stephen Obinna; with appendix by: Dan Abramovich, Veronica Arena, Stephen Obinna

The inverse Toeplitz eigenvalue problem (ToIEP) concerns finding a vector that specifies the real-valued symmetric Toeplitz matrix with the prescribed set of eigenvalues. Since phase "calibration" errors in uniform linear antenna arrays (ULAs) do not change the covariance matrix eigenvalues and the moduli of the covariance matrix elements, we formulate a number of the new ToIEP problems of the Hermitian Toeplitz matrix reconstruction, given the moduli of the matrix elements and the matrix eigenvalues. We demonstrate that for the real-valued case, only two solutions to this problem exist, with the "non-physical" one that in most practical cases could be easily disregarded. The computational algorithm for the real-valued case is quite simple. For the complex-valued case, we demonstrate that the family of solutions is broader and includes solutions inappropriate for calibration. For this reason, we modified this ToIEP problem to match the covariance matrix of the uncalibrated ULA. We investigate the statistical convergence of the ad-hoc algorithm with the sample matrices instead of the true ones. The proposed ad-hoc algorithms require the so-called "strong" or "argumental" convergence, which means a large enough required sample volume that reduces the errors in the estimated covariance matrix elements. Along with the ULA arrays, we also considered the fully augmentable minimum redundancy arrays that generate the same (full) set of covariance lags as the uniform linear arrays, and we specified the conditions when the ULA Toeplitz covariance matrix may be reconstructed given the M-variate MRA covariance matrix., Comment: 27 pages, 40 figures

The problem of a constant modulus (CM) continuous wave (CW) waveform design with the thumb-tack ambiguity function that meets the NTIA RSEC requirements is addressed. The ad-hoc and alternating projection techniques are proposed to modify the spectrum of a prototype waveform to meet the NTIA RSEC requirements, retaining the "thumb-tack" property of the original waveform ambiguity function. Periodic binary frequency shift keying (BFSK), Costas FSK, and noise-like waveforms are modified to meet CM and NTIA RSEC requirements. Introduced examples demonstrate the spectrum modification consequences and the proposed technique's efficiency for generating the CM spectrum-controlled waveforms with the thumb-tack ambiguity function.

Action recognition models have achieved impressive results by incorporating scene-level annotations, such as objects, their relations, 3D structure, and more. However, obtaining annotations of scene structure for videos requires a significant amount of effort to gather and annotate, making these methods expensive to train. In contrast, synthetic datasets generated by graphics engines provide powerful alternatives for generating scene-level annotations across multiple tasks. In this work, we propose an approach to leverage synthetic scene data for improving video understanding. We present a multi-task prompt learning approach for video transformers, where a shared video transformer backbone is enhanced by a small set of specialized parameters for each task. Specifically, we add a set of "task prompts", each corresponding to a different task, and let each prompt predict task-related annotations. This design allows the model to capture information shared among synthetic scene tasks as well as information shared between synthetic scene tasks and a real video downstream task throughout the entire network. We refer to this approach as "Promptonomy", since the prompts model task-related structure. We propose the PromptonomyViT model (PViT), a video transformer that incorporates various types of scene-level information from synthetic data using the "Promptonomy" approach. PViT shows strong performance improvements on multiple video understanding tasks and datasets. Project page: \url{https://ofir1080.github.io/PromptonomyViT}, Comment: WACV 2024

We consider (nonparametric) sparse (generalized) additive models (SpAM) for classification. The design of a SpAM classifier is based on minimizing the logistic loss with a sparse group Lasso/Slope-type penalties on the coefficients of univariate additive components' expansions in orthonormal series (e.g., Fourier or wavelets). The resulting classifier is inherently adaptive to the unknown sparsity and smoothness. We show that under certain sparse group restricted eigenvalue condition it is nearly-minimax (up to log-factors) simultaneously across the entire range of analytic, Sobolev and Besov classes. The performance of the proposed classifier is illustrated on a simulated and a real-data examples.

We show how the notion of fantastacks can be used to effectively desingularize binomial varieties defined over algebraically closed fields. In contrast to a desingularization via blow-ups in smooth centers, we drastically reduce the number of steps and the number of charts appearing along the process. Furthermore, we discuss how our considerations extend to a partial simultaneous normal crossings desingularization of finitely many binomial hypersurfaces., Comment: Minor changes, accepted for publication in the Mathematische Zeitschrift, 23 pages

Objective: Ophthalmological pathologies such as glaucoma, diabetic retinopathy and age-related macular degeneration are major causes of blindness and vision impairment. There is a need for novel decision support tools that can simplify and speed up the diagnosis of these pathologies. A key step in this process is to automatically estimate the quality of the fundus images to make sure these are interpretable by a human operator or a machine learning model. We present a novel fundus image quality scale and deep learning (DL) model that can estimate fundus image quality relative to this new scale. Methods: A total of 1,245 images were graded for quality by two ophthalmologists within the range 1-10, with a resolution of 0.5. A DL regression model was trained for fundus image quality assessment. The architecture used was Inception-V3. The model was developed using a total of 89,947 images from 6 databases, of which 1,245 were labeled by the specialists and the remaining 88,702 images were used for pre-training and semi-supervised learning. The final DL model was evaluated on an internal test set (n=209) as well as an external test set (n=194). Results: The final DL model, denoted FundusQ-Net, achieved a mean absolute error of 0.61 (0.54-0.68) on the internal test set. When evaluated as a binary classification model on the public DRIMDB database as an external test set the model obtained an accuracy of 99%. Significance: the proposed algorithm provides a new robust tool for automated quality grading of fundus images., Comment: 12 pages, 9 figures, published in Computer Methods and Programs in Biomedicine

We consider high-dimensional multiclass classification by sparse multinomial logistic regression. Unlike binary classification, in the multiclass setup one can think about an entire spectrum of possible notions of sparsity associated with different structural assumptions on the regression coefficients matrix. We propose a computationally feasible feature selection procedure based on penalized maximum likelihood with convex penalties capturing a specific type of sparsity at hand. In particular, we consider global sparsity, double row-wise sparsity, and low-rank sparsity, and show that with the properly chosen tuning parameters the derived plug-in classifiers attain the minimax generalization error bounds (in terms of misclassification excess risk) within the corresponding classes of multiclass sparse linear classifiers. The developed approach is general and can be adapted to other types of sparsity as well.