Multimodal Large Language Models (MLLMs) largely lag human-level performance on abstract visual reasoning (AVR), which requires models to infer latent rules from visual question sets and generalize them to novel scenarios. Most AVR benchmarks are constrained to narrow and repetitive 2D patterns, involving relatively simple spatial relationships and assessing limited dimensions of reasoning ability. Drawing inspiration from real-world paper folding challenges, we propose Paper Folding Puzzles (PFP), a rigorously designed benchmark specifically developed to assess spatial reasoning capabilities. It comprises 150K visual question-answering samples across five diverse tasks, ranging from basic 2D geometric reasoning to 3D spatial understanding. The developed benchmark dataset can be employed to assess core spatial reasoning abilities essential to human cognition, encompassing fundamental symmetry reasoning and 3D spatial comprehension. Furthermore, we conduct a comprehensive evaluation of 18 leading MLLMs (both closed- and open-source variants) on the PFP benchmark to assess their spatial reasoning capabilities. Our findings show that most MLLMs achieve near-chance performance on FPF, exhibiting substantial performance gaps (>30%) relative to human baselines across all tasks. This highlights a critical research gap in improving spatial reasoning capabilities of MLLMs.

ABSTRACT Interactive computer technology is deeply integrated into traditional teaching methods. The traditional teaching of physics experiments in secondary schools suffers from the inability of teachers to provide timely guidance to students, the difficulty of controlling experimental variables, and the lack of uniformity in evaluation criteria. To address the aforementioned issues, we have developed an innovative system to improve secondary school physics education using computer vision‐based interaction with virtual humans and sensors. The proposed system captures experimental data in real time so that student performance can be accurately monitored and assessed. Teachers can effortlessly configure experiments through simple coding, while the system leverages a multimodal macrolanguage model to offer contextual feedback and guidance. The system generates a virtual teacher that offers step‐by‐step guidance and real‐time feedback. Usability tests indicate that the system significantly improves student engagement and comprehension of complex physics concepts, highlighting its potential to transform traditional science education. The advantage of real‐time assessment and guidance in secondary school physics experiments is that it enables students to grasp abstract concepts in a more intuitive and comprehensible manner.

Given sparse point clouds, this paper develops a perturbation learning-based point cloud upsampling method to generate uniform, clean, and dense point clouds. We build a simple yet efficient neural network framework including feature extraction, perturbation learning, and coordinate reconstruction operations. In the feature extraction task, shallow-and-wide dense connections are applied to present the latent geometric information. Subsequently, the extracted features are expanded for perturbation learning. According to the theory of the differential geometry of surfaces, the position of an upsampled point can be approximated by its projection on a tangent plane in a sufficiently small neighborhood around the point. Inspired by this, we propose learning a 2D perturbation through multilayer perceptrons (MLPs) to estimate the coordinate shift from the input point to the upsampled point. Then, we concatenate the 2D perturbation with the extracted features for residual learning to fine-tune the coordinate shift. Finally, the coordinate reconstruction step transforms all the high-level features into an upsampled and consolidated point cloud. To enable the learning-based point cloud upsampling process above, we collect a large-scale point cloud dataset that contains 36000 pairs of training sets. The entire network size is only 5.01 MB, which is much smaller than the requirements of state-of-the-art point cloud upsampling models. Qualitative and quantitative evaluation results show that our proposed scheme outperforms most existing methods in terms of the Chamfer distance (CD), Hausdorff distance (HD), Jensen-Shannon divergence (JSD), and uniformity. In addition, our method is applied to real scan data, and its robustness is confirmed.

State-of-the-art Convolutional Neural Network (CNN) based loop restoration generally involves a CNN structure with a large number of parameters and applies the CNN model to those degraded frames uniformly to generate their restored version, even though the contents within these frames are different. By contrast, in this paper, we propose a Guided CNN Restoration (GNR) scheme, where a CNN is used in conjunction with explicitly signaled guide parameters, with an aim to adapt the CNN model to different input contents. Specifically, the CNN architecture is designed such that the final restoration is constrained within the subspace generated by various output channels of the CNN, and meanwhile the weighting parameters for a linear combination of the output channels to obtain the final restoration are explicitly signaled by the encoders. The proposed GNR is incorporated into an AV1 encoder to replace the anchor in-loop filters and the weighting parameters are written into the encoded bitstream. Experimental results show that given a small CNN with 3,312 parameters, the proposed approach achieves a BD-rate reduction of 3.06% over the AV1 anchor, while the traditional CNN-based method only achieves 1.39%.

Most techniques for real-time construction of a signed distance field, whether on a CPU or GPU, involve approximate distances. We use a GPU to build an exact adaptive distance field, constructed from an octree by using the Morton code. We use rectangle-swept spheres to construct a bounding volume hierarchy (BVH) around a triangulated model. To speed up BVH construction, we can use a multi-BVH structure to improve the workload balance between GPU processors. An upper bound on distance to the model provided by the octree itself allows us to reduce the number of BVHs involved in determining the distances from the centers of octree nodes at successively lower levels, prior to an exact distance query involving the remaining BVHs. Distance fields can be constructed 35-64 times as fast as a serial CPU implementation of a similar algorithm, allowing us to simulate a piece of fabric interacting with the Stanford Bunny at 20 frames per second.

We cull collisions between very large numbers of moving bodies using graphics processing units (GPUs). To perform massively parallel sweep-and-prune (SaP), we mitigate the great density of intervals along the axis of sweep by using principal component analysis to choose the best sweep direction, together with spatial subdivisions to further reduce the number of false positive overlaps. Our algorithm implemented entirely on GPUs using the CUDA framework can handle a million moving objects at interactive rates. As application of our algorithm, we demonstrate the real-time simulation of very large numbers of particles and rigid-body dynamics.