Large-scale Vision-Language Models (LVLMs) process both images and text, excelling in multimodal tasks such as image captioning and description generation. However, while these models excel at generating factual content, their ability to generate and evaluate texts reflecting perspectives on the same image, depending on the context, has not been sufficiently explored. To address this, we propose IRR: Image Review Rank, a novel evaluation framework designed to assess critic review texts from multiple perspectives. IRR evaluates LVLMs by measuring how closely their judgments align with human interpretations. We validate it using a dataset of images from 15 categories, each with five critic review texts and annotated rankings in both English and Japanese, totaling over 2,000 data instances. Our results indicate that, although LVLMs exhibited consistent performance across languages, their correlation with human annotations was insufficient, highlighting the need for further advancements. These findings highlight the limitations of current evaluation methods and the need for approaches that better capture human reasoning in Vision & Language tasks.

Hongyu Sun, Yusuke Sakai, Haruki Sakajo, Shintaro Ozaki, Kazuki Hayashi, Hidetaka Kamigaito, Taro Watanabe. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Spectral geometry is a mathematical field that links geometrical properties to eigenvalues of differential operators on surfaces. Although it is a well-established tool in geometry processing and has been used in many contexts, the structural engineering and architectural geometry communities have not yet adopted this framework for shape modeling. This paper aims to explore spectral methods for applications in architectural geometries. A novel methodology for generating anisotropic Laplacian operators based on regions of interest defined by the user is proposed. The potential of spectral methods in structural design is illustrated through design problems expressed on meshes and graphs.

Piecewise constant mean curvature (P-CMC) surfaces are generated using the mean curvature flow (MCF). As an extension of the known fact that a CMC surface is the stationary point of an energy functional, a P-CMC surface can be obtained as the stationary point of an energy functional of multiple patch surfaces and auxiliary surfaces between them. A new formulation is presented for the MCF as the negative gradient flow of the energy functional for multiple patch continuous surfaces, which are further discretized so as to determine the change in the vertex positions of triangular meshes on the surface as well as along the internal boundaries between patches. Numerical examples show that multiple patch surfaces approximately reach the specified mean curvatures through the proposed method, which can diversify the options for the shape design using CMC surfaces.

A combined method of graph embedding (GE) and reinforcement learning (RL) is developed for discrete cross-section optimization of planar steel frames, in which the section size of each member is selected from a prescribed list of standard sections. The RL agent aims to minimize the total structural volume under various practical constraints. GE is a method for extracting features from data with irregular connectivity. While most of the existing GE methods aim at extracting node features, an improved GE formulation is developed for extracting features of edges associated with members in this study. Owing to the proposed GE operations, the agent is capable of grasping the structural property of columns and beams considering their connectivity in a frame with an arbitrary size as feature vectors of the same size. Using the feature vectors, the agent is trained to estimate the accurate return associated with each action and to take proper actions on which members to reduce or increase their size using an RL algorithm. The applicability of the proposed method is versatile because various frames different in the numbers of nodes and members can be used for both training and application phases. In the numerical examples, the trained agents outperform a particle swarm optimization method as a benchmark in terms of both computational cost and design quality for cross-sectional design changes; the agents successfully assign reasonable cross-sections considering the geometry, connectivity, and support and load conditions of the frames.

A method is presented for generating a discrete piecewise constant Gaussian curvature (CGC) surface. An energy functional is first formulated so that its stationary point is the linear Weingarten (LW) surface, which has a property such that the weighted sum of mean and Gaussian curvatures is constant. The CGC surface is obtained using the gradient derived from the first variation of a special type of the energy functional of the LW surface and updating the surface shape based on the Gaussian curvature flow. A filtering method is incorporated to prevent oscillation and divergence due to unstable property of the discretized Gaussian curvature flow. Two techniques are proposed to generate a discrete piecewise CGC surface with preassigned internal boundaries. The step length of Gaussian curvature flow is adjusted by introducing a line search algorithm to minimize the energy functional. The effectiveness of the proposed method is demonstrated through numerical examples of generating various shapes of CGC surfaces.

To plan behavior procedures, it is necessary for an agent to have a world model concerning the temporal sequences information. In this paper, a temporal information learning algorithm is proposed with a three layer neural network implementing the "effectiveness of simulation accumulation" algorithm. This algorithm can construct a "behavior automaton" in the neural network. From the results of some learning experiments using a mobile robot simulation, the generated automaton expresses the complexity of the simulation environments. The robot agent acquires a behavior automaton for obstacle avoidance behavior which is influenced by the simulation environment.