Large language models (LLMs) are increasingly applied in clinical decision support, yet current evaluations rarely reveal whether their outputs reflect genuine medical reasoning or superficial correlations. We introduce DeVisE (Demographics and Vital signs Evaluation), a behavioral testing framework that probes fine-grained clinical understanding through controlled counterfactuals. Using intensive care unit (ICU) discharge notes from MIMIC-IV, we construct both raw (real-world) and template-based (synthetic) variants with single-variable perturbations in demographic (age, gender, ethnicity) and vital sign attributes. We evaluate eight LLMs, spanning general-purpose and medical variants, under zero-shot setting. Model behavior is analyzed through (1) input-level sensitivity, capturing how counterfactuals alter perplexity, and (2) downstream reasoning, measuring their effect on predicted ICU length-of-stay and mortality. Overall, our results show that standard task metrics obscure clinically relevant differences in model behavior, with models differing substantially in how consistently and proportionally they adjust predictions to counterfactual perturbations.
In this work, we present a methodology for devising forward-backward methods for finding zeros in the sum of a finite number of maximally monotone operators. We extend the framework and techniques from [SIAM J. Optim., 34 (2024), pp. 1569-1594] to cover the case involving a finite number of cocoercive operators, which should be directly evaluated instead of computing their resolvent. The algorithms are induced by three graphs that determine how the algorithm variables interact with each other and how they are combined to compute each resolvent. The hypotheses on these graphs ensure that the algorithms obtained have minimal lifting and are frugal, meaning that the ambient space of the underlying fixed point operator has minimal dimension and that each resolvent and each cocoercive operator is evaluated only once per iteration. This framework not only allows to recover some known methods, but also to generate new ones, as the forward-backward algorithm induced by a complete graph. We conclude with a numerical experiment showing how the choice of graphs influences the performance of the algorithms.
Spiking Neural Networks (SNNs) emulate the spiking behavior of biological neurons and are typically deployed on distributed-memory neuromorphic hardware. The deployment of a SNN usually requires partitioning the network and mapping these partitions onto the hardware's processing units. However, finding optimal deployment configurations is an NP-hard problem, often addressed through optimization algorithms. While some objectives (e.g., memory utilization and chip count) are static, others (e.g., communication latency and energy efficiency) depend on the network's dynamic behavior, necessitating dynamic-aware optimization. To address this, we model SNN dynamics using an Ising-type pairwise interaction framework, bridging microscopic neuron interactions with macroscopic network behavior. We optimize deployment by exploring the parameter and configuration spaces of the Ising model. We evaluate our approach on two SNNs deployed on the sPyNNaker neuromorphic platform. Initial results suggest that the method underperforms, potentially due to the Ising model's equilibrium assumptions and the architectural complexity of real-world neuromorphic hardware, highlighting limitations in its curre
We study the $(Δ+1)$-edge-coloring problem in the parallel $\left(\mathrm{PRAM}\right)$ model of computation. The celebrated Vizing's theorem [Viz64] states that every simple graph $G = (V,E)$ can be properly $(Δ+1)$-edge-colored. In a seminal paper, Karloff and Shmoys [KS87] devised a parallel algorithm with time $O\left(Δ^5\cdot\log n\cdot\left(\log^3 n+Δ^2\right)\right)$ and $O(m\cdotΔ)$ processors. This result was improved by Liang et al. [LSH96] to time $O\left(Δ^{4.5}\cdot \log^3Δ\cdot \log n + Δ^4 \cdot\log^4 n\right)$ and $O\left(n\cdotΔ^{3} +n^2\right)$ processors. [LSH96] claimed $O\left(Δ^{3.5} \cdot\log^3Δ\cdot \log n + Δ^3\cdot \log^4 n\right)$ time, but we point out a flaw in their analysis, which once corrected, results in the above bound. We devise a faster parallel algorithm for this fundamental problem. Specifically, our algorithm uses $O\left(Δ^4\cdot \log^4 n\right)$ time and $O(m\cdot Δ)$ processors. Another variant of our algorithm requires $O\left(Δ^{4+o(1)}\cdot\log^2 n\right)$ time, and $O\left(m\cdotΔ\cdot\log n\cdot\log^δΔ\right)$ processors, for an arbitrarily small $δ>0$. We also devise a few other tradeoffs between the time and the number of process
We present a rehearsal-oriented system for creating responsive built environments during theatre devising workshops. The system connects bespoke sensing modules for gesture, position, and speech recognition to light and sound outputs through a visual no-code programming layer. It was developed, used, and refined across six workshops with eight professional performance-makers, where participants created light-and-sound scores, gesture- and position-triggered scenes, responsive architectures, participatory prototypes, and a multi-room scratch performance. Rather than presenting a production-ready show-control platform, this demo focuses on how sensing and actuation can be made available as compositional materials during early-stage creative experimentation for immersive theatrical compositions. The system is designed to support quick configuration, visible mappings, and in-room testing, allowing performers to experiment with responsive spaces with minimal technical support. We describe the system architecture, its workshop use, and the practical conditions that helped integrate interactive sensing into embodied performance-making.
Given a graph $G=(V,E)$, a collection $\mathcal{T}$ of spanning trees of $G$ is called a spanning tree cover of stretch $α$ if for every $u,v\in V$ there is a tree $T_{uv}\in\mathcal{T}$, such that \[d_{T_{uv}}(u,v)\leqα\cdot d_G(u,v)~.\] Spanning tree covers were introduced in the pioneering work of Gupta et al. [GKR04], that showed that $p$-path-separable graphs admit stretch-$3$ spanning tree covers with size $O(p\log n)$. Many subsequent papers focused on a relaxed notion of non-spanning tree covers, in which the trees are required to be dominating, but may use edges that do not belong to the graph. In particular, Bartal et al. [BFN22] devised a construction of non-spanning tree covers with stretch $1+ε$ and size $O(p\cdot\frac{\log^2n}{ε^2})$. Recently, for $K_r$-minor-free graphs, Chang et al. [CCL+23,CCL+24] devised a non-spanning tree cover with stretch $1+ε$ and size $2^{\frac{1}εr^{O(r)}}$, and an exact spanning tree cover with size $r^{O(diam(G))}$. However, the problem of devising spanning tree covers with stretch smaller than $3$ and small size for general $p$-path-separable graphs remained open. We show that $p$-path-separable graphs admit spanning tree covers with st
Denoising diffusion models produce high-fidelity image samples by capturing the image distribution in a progressive manner while initializing with a simple distribution and compounding the distribution complexity. Although these models have unlocked new applicabilities, the sampling mechanism of diffusion does not offer means to extract image-specific semantic representation, which is inherently provided by auto-encoders. The encoding component of auto-encoders enables mapping between a specific image and its latent space, thereby offering explicit means of enforcing structures in the latent space. By integrating an encoder with the diffusion model, we establish an auto-encoding formulation, which learns image-specific representations and offers means to organize the latent space. In this work, First, we devise a mechanism to structure the latent space of a diffusion auto-encoding model, towards recognizing region-specific cellular patterns in brain images. We enforce the representations to regress positional information of the patches from high-resolution images. This creates a conducive latent space for differentiating tissue types of the brain. Second, we devise an unsupervised
In this work, to facilitate the real-time processing of multi-scan registration error minimization on factor graphs, we devise a point cloud downsampling algorithm based on coreset extraction. This algorithm extracts a subset of the residuals of input points such that the subset yields exactly the same quadratic error function as that of the original set for a given pose. This enables a significant reduction in the number of residuals to be evaluated without approximation errors at the sampling point. Using this algorithm, we devise a complete SLAM framework that consists of odometry estimation based on sliding window optimization and global trajectory optimization based on registration error minimization over the entire map, both of which can run in real time on a standard CPU. The experimental results demonstrate that the proposed framework outperforms state-of-the-art CPU-based SLAM frameworks without the use of GPU acceleration.
We devise an a posteriori error estimator for an affine optimal control problem subject to a semilinear elliptic PDE and control constraints. To approximate the problem, we consider a semidiscrete scheme based on the variational discretization approach. For this solution technique, we design an a posteriori error estimator that accounts for the discretization of the state and adjoint equations, and prove, under suitable local growth conditions of optimal controls, reliability and efficiency properties of such error estimator. A simple adaptive strategy based on the devised estimator is designed and its performance is illustrated with numerical examples.
Constant workspace algorithms use a constant number of words in addition to the read-only input to the algorithm. In this paper, we devise algorithms to efficiently compute relative hulls in the plane using a constant workspace. Specifically, we devise algorithms for the following three problems: (i) Given two simple polygons P and Q with P \subset Q, compute a simple polygon P' with a perimeter of minimum length such that P \subseteq P' \subseteq Q. (ii) Given two simple polygons P and Q such that Q does not intersect the relative interior of P but it does intersect the relative interior of the convex hull of P, compute a weakly simple polygon P' with a perimeter of minimum length such that P \subseteq P', the convex hull of P contains P', and P' does not intersect the relative interior of Q. (iii) Given a set S of points located in a simple polygon P, compute a weakly simple polygon P' with a perimeter of minimum length such that P' \subseteq P and P' contains all the points in S. To our knowledge, no prior work devised algorithms to compute relative hulls using a constant workspace, and this work is the first such attempt.
Fuzzing has been proven extremely effective in finding vulnerabilities in software. When it comes to fuzz stateless systems, analysts have no doubts about the choice to make. In fact, among the plethora of stateless fuzzers devised in the last 20 years, AFL (with its descendants AFL++ and LibAFL) stood up for its effectiveness, speed and ability to find bugs. On the other hand, when dealing with stateful systems, it is not clear what is the best tool to use. In fact, the research community struggles to devise (and benchmark) effective and generic stateful fuzzers. In this short paper, we discuss the reasons that make stateful fuzzers difficult to devise and benchmark.
We consider a control-constrained optimal control problem subject to time-harmonic Maxwell's equations; the control variable belongs to a finite-dimensional set and enters the state equation as a coefficient. We derive existence of optimal solutions, and analyze first- and second-order optimality conditions. We devise an approximation scheme based on the lowest order Nédélec finite elements to approximate optimal solutions. We analyze convergence properties of the proposed scheme and prove a priori error estimates. We also design an a posteriori error estimator that can be decomposed as the sum two contributions related to the discretization of the state and adjoint equations, and prove that the devised error estimator is reliable and locally efficient. We perform numerical tests in order to assess the performance of the devised discretization strategy and the a posteriori error estimator.
According to the data, the percent of women who underwent screening for cervical cancer, breast and oral cancer in Telangana in the year 2020 was 3.3 percent, 0.3 percent and 2.3 percent respectively. Although early detection is the only way to reduce morbidity and mortality, people have very low awareness about cervical and breast cancer signs and symptoms and screening practices. We developed an ML classification model to predict if a person is susceptible to breast or cervical cancer based on demographic factors. We devised a system to provide suggestions for the nearest hospital or Cancer treatment centres based on the users location or address. In addition to this, we can integrate the health card to maintain medical records of all individuals and conduct awareness drives and campaigns. For ML classification models, we used decision tree classification and support vector classification algorithms for cervical cancer susceptibility and breast cancer susceptibility respectively. Thus, by devising this solution we come one step closer to our goal which is spreading cancer awareness, thereby, decreasing the cancer mortality and increasing cancer literacy among the people of Telang
In recent years, we have observed a growing interest in Indoor Tracking Systems (ITS) for providing location-based services indoors. This is due to the limitations of Global Navigation and Satellite Systems, which do not operate in non-line-of-sight environments. Depending on their architecture, ITS can rely on expensive infrastructure, accumulate errors, or be challenging to evaluate in real-life environments. Building an ITS is a complex process that involves devising, evaluating and fine-tuning tracking algorithms. This process is not yet standard, as researchers use different types of equipment, deployment environments, and evaluation metrics. Therefore, it is challenging for researchers to build novel tracking algorithms and for the research community to reproduce the experiments. To address these challenges, we propose MobiXIM, a framework that provides a set of tools for devising, evaluating and fine-tuning tracking algorithms in a structured manner. For devising tracking algorithms, MobiXIM introduces a novel plugin architecture, allowing researchers to collaborate and extend existing algorithms. We assess our framework by building an ITS encompassing the key elements of wi
In digital pathology, precise nuclei segmentation is pivotal yet challenged by the diversity of tissue types, staining protocols, and imaging conditions. Recently, the segment anything model (SAM) revealed overwhelming performance in natural scenarios and impressive adaptation to medical imaging. Despite these advantages, the reliance of labor-intensive manual annotation as segmentation prompts severely hinders their clinical applicability, especially for nuclei image analysis containing massive cells where dense manual prompts are impractical. To overcome the limitations of current SAM methods while retaining the advantages, we propose the Universal prompt-free SAM framework for Nuclei segmentation (UN-SAM), by providing a fully automated solution with remarkable generalization capabilities. Specifically, to eliminate the labor-intensive requirement of per-nuclei annotations for prompt, we devise a multi-scale Self-Prompt Generation (SPGen) module to revolutionize clinical workflow by automatically generating high-quality mask hints to guide the segmentation tasks. Moreover, to unleash the generalization capability of SAM across a variety of nuclei images, we devise a Domain-adapt
This paper reports a practice-based investigation into authoring responsive light and sound in immersive performance without writing code. A modular system couples live gesture, position, and speech inputs to scenographic outputs through a visual logic layer that performers can operate in rehearsal. Across six workshops with eight professional performance-makers, we staged a progression from parallel ensemble and technical training to integrated dramaturgy, culminating in a single-spectator scratch immersive performance with interactive elements. This paper details the system's building blocks and the workshop arc. A reflexive reading of workshop video logs, post-workshop focus groups, and facilitator notes surfaced three ensemble-level strategies that made the technology workable in a hybrid devising/design practice: rotating roles between operator, performer, and mediator; embracing controlled imperfection as a creative resource; and using technology-describing metaphors to support creative practice.
Large language models (LLMs) remain vulnerable to a slew of adversarial attacks and jailbreaking methods. One common approach employed by white-hat attackers, or red-teamers, is to process model inputs and outputs using string-level obfuscations, which can include leetspeak, rotary ciphers, Base64, ASCII, and more. Our work extends these encoding-based attacks by unifying them in a framework of invertible string transformations. With invertibility, we can devise arbitrary string compositions, defined as sequences of transformations, that we can encode and decode end-to-end programmatically. We devise a automated best-of-n attack that samples from a combinatorially large number of string compositions. Our jailbreaks obtain competitive attack success rates on several leading frontier models when evaluated on HarmBench, highlighting that encoding-based attacks remain a persistent vulnerability even in advanced LLMs.
Information disclosure can compromise privacy when revealed information is correlated with private information. We consider the notion of inferential privacy, which measures privacy leakage by bounding the inferential power a Bayesian adversary can gain by observing a released signal. Our goal is to devise an inferentially-private private information structure that maximizes the informativeness of the released signal, following the Blackwell ordering principle, while adhering to inferential privacy constraints. To achieve this, we devise an efficient release mechanism that achieves the inferentially-private Blackwell optimal private information structure for the setting where the private information is binary. Additionally, we propose a programming approach to compute the optimal structure for general cases given the utility function. The design of our mechanisms builds on our geometric characterization of the Blackwell-optimal disclosure mechanisms under privacy constraints, which may be of independent interest.
The linear Fisher market (LFM) is a basic equilibrium model from economics, which also has applications in fair and efficient resource allocation. First-price pacing equilibrium (FPPE) is a model capturing budget-management mechanisms in first-price auctions. In certain practical settings such as advertising auctions, there is an interest in performing statistical inference over these models. A popular methodology for general statistical inference is the bootstrap procedure. Yet, for LFM and FPPE there is no existing theory for the valid application of bootstrap procedures. In this paper, we introduce and devise several statistically valid bootstrap inference procedures for LFM and FPPE. The most challenging part is to bootstrap general FPPE, which reduces to bootstrapping constrained M-estimators, a largely unexplored problem. We devise a bootstrap procedure for FPPE under mild degeneracy conditions by using the powerful tool of epi-convergence theory. Experiments with synthetic and semi-real data verify our theory.
Alzheimer's disease (AD) has become one of the most significant health challenges in an aging society. The use of spoken language-based AD detection methods has gained prevalence due to their scalability due to their scalability. Based on the Cookie Theft picture description task, we devised an explainable and effective feature set that leverages the visual capabilities of a large language model (LLM) and the Term Frequency-Inverse Document Frequency (TF-IDF) model. Our experimental results show that the newly proposed features consistently outperform traditional linguistic features across two different classifiers with high dimension efficiency. Our new features can be well explained and interpreted step by step which enhance the interpretability of automatic AD screening.