Hyperbolic vision-language models are designed to encode abstraction geometrically: general concepts near the origin, specific ones farther out, and entailment cones representing directed order. We ask whether trained MERU, HyCoCLIP, and PHyCLIP models actually use these mechanisms. We audit seven released checkpoints and matched from-scratch interventions. All audited converged checkpoints remain near-Euclidean in the dimensionless radius u=√(c)ρ: the largest observed image-side value is 0.37 – well below u≈0.84, where local metric distortion reaches 10%. Releasing the curvature floor changes curvature and norms but not this regime, with mixed, generally modest downstream shifts. Trained entailment cones are saturated or nearly saturated, so low violation rates can arise from trivially wide cones rather than learned order. Preregistered semantic traversal detects weak within-branch order but no operative full-hierarchy readout. Shuffle-controlled tests detect no pair-specific radial ordering in released checkpoints, and no positive result is consistent across all three matched ViT-B seeds. This non-detection is sensitivity-bounded: the diagnostic reaches 80% power only at ≈13.5 pp of shuffle-controlled excess, while the largest released gap is ≈5 pp. We trace this to a low-curvature shortcut: lowering curvature widens entailment cones and suppresses violations without learning order. In the probed trajectories, gradient decomposition identifies entailment as the dominant curvature-lowering pressure during collapse. Yet curvature contracts even when entailment is removed, so the shortcut is not the sole cause. Under our diagnostics, the audited formulations do not demonstrate an operative radial or cone-based hierarchy. We distill the audit into a five-number geometry report for evaluating future hierarchy claims.
A new mathematical mesomodel of object flows for transportation problems is presented. The mathematical model is based on the system of Vlasov–Poisson equations. The model describes the motion of individual objects, taking into account their interaction through the formed information field, which provides a collision limitation. The problem of motion of a set of interacting objects in a narrow sector is considered. It is proposed to use the PIC method. The method consists in solving the equations of motion of objects, which correspond to the characteristics of Vlasov’s kinetic equations. The equations of motion take into account the information tension and a set of additional flow factors. This will allow the model to be used to describe land, air and water transport flows.
One of the world's greatest challenges is the effort to reduce greenhouse gas emissions. One of the major sources of these emissions is road transport. The government is introducing a number of restrictions on the operation of internal combustion engine vehicles, such as a ban on warming up the engine. At the same time, car manufacturers are developing start–stop systems. All this is aimed at reducing fuel combustion and atmospheric emissions. Despite the complete phase-out of internal combustion engines by 2050, such vehicles will still form a significant part of the European car fleet. On the other hand, one of the reasons for the slow distribution of electric vehicles is their limited range. Both of these problems can be mitigated by energy-efficient driving. And how does the infrastructure of a city affect car operation and consequently fuel burn? The paper presents a mathematical model for energy-efficient control of connected highly automated vehicles in the conditions of the modern metropolis. The mathematical model incorporates terrain, speed modes, and traffic light schedules. Laboratory tests of the developed mathematical model are carried out on a digital twin of a test road section recreated on the basis of high-precision navigation data and a passenger vehicle with an internal combustion engine. In the laboratory tests, the proposed system is compared with a vehicle being driven on cruise control. The average speed of both vehicles on the selected route is identical. The results of the experimental runs indicate that the proposed mathematical model is 4.5
В статье исследован метод хранения изображений путем обучения нейронной сети на одном изображении и сохранения ее весов как компактного представления. Данный подход позволяет значительно уменьшить объем хранимых данных при сохранении приемлемого визуального качества. Проанализированы параметры модели и настройки обучения с целью оптимизации качества восстановления. Основная идея подхода заключается в том, что обученная модель сохраняет аппроксимированное отображение изображения в свои веса, которые выступают в качестве компактного представления исходного изображения. При необходимости реконструкции веса загружаются обратно в сеть для восстановления визуального содержания. Экспериментальные результаты показывают, что оптимизация архитектуры сети и цветового пространства (YCbCr) позволяет достичь высоких коэффициентов сжатия до 29,4 при сохранении визуального качества, близкого к оригиналу (MSE ≈ 10-5). Однако авторы отмечают существенный недостаток метода: длительное время обучения и значительные вычислительные затраты, что делает его менее эффективным по сравнению с традиционными алгоритмами сжатия для практического применения в реальном времени. Тем не менее, подход демонстрирует потенциал для задач, где критически важно сохранение тонких деталей изображений, например, архивации данных или сжатия видеопотоков. This article explores a method for storing images by training a neural network on a single image and storing its weights as a compact representation. This approach significantly reduces the amount of data stored while maintaining acceptable visual quality. Model parameters and training settings are analyzed to optimize recovery quality. The basic idea of the approach is that a trained model stores its weights, which act as a compact representation of the original image. When reconstruction is required, the weights are reloaded into the network to restore the visual content. Experimental results show that optimizing the network architecture and color space (YCbCr) enables high compression ratios – up to 29.4 while maintaining visual quality close to the original (MSE ≈ 10-5). However, the authors note a significant drawback of the method: long training time and significant computational costs, making it less effective than traditional compression algorithms for practical real-time applications. Nevertheless, the approach demonstrates potential for tasks where preserving fine image details is critical, such as data archiving or video stream compression.
The study of rheokinetic characteristics is one of the most important stages in the development of new formulations of modified polymeric materials. Examining the specifics of the curing process of compositions makes it possible to assess component compatibility and the influence of various additives on a set of technological and performance properties. When developing polymer composite materials modeled after natural layered materials with weak interfaces (wood, shells, certain types of mica, etc.), an important technological task is to preserve, by means of applied modifiers, an original liquid aggregate state with the formation of an independent liquid phase within the composite structure (in this work, these modifiers are referred to as liquid-phase materials; triethylene glycol dimethacrylate and a silicone sealant are used as examples). The role of liquid-phase materials, arranged in the composite structure according to specified reinforcement schemes, is to avoid defect accumulation under deformation and to provide stress relaxation. Therefore, when designing a polymer composite composition with a two-phase reinforcement scheme, one of the key criteria for selecting a liquid-phase material is the absence of chemical interaction with the binder material. In this paper, using rotational viscometry, the influence of triethylene glycol dimethacrylate and a silicone sealant on changes in the rheokinetic characteristics of the ED-20 epoxy binder with the amine hardener PEPA is studied at +40, +60, and +80°C. It is found that adding these materials to the epoxy binder in the amount of 20 parts by mass per 100 parts by mass of the epoxy composition does not lead to a significant change in gelation time, viscosity growth constants, or activation energies, which indicates the absence of chemical interaction between the components.