TangramPuzzle: Evaluating Multimodal Large Language Models with Compositional Spatial Reasoning

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in visual recognition and semantic understanding. Nevertheless, their ability to perform precise compositional spatial reasoning remains largely unexplored. Existing benchmarks often involve relatively simple tasks and rely on semantic approximations or coarse relative positioning, while their evaluation metrics are typically limited and lack rigorous mathematical formulations. To bridge this gap, we introduce TangramPuzzle, a geometry-grounded benchmark designed to evaluate compositional spatial reasoning through the lens of the classic Tangram game. We propose the Tangram Construction Expression (TCE), a symbolic geometric framework that grounds tangram assemblies in exact, machine-verifiable coordinate specifications, to mitigate the ambiguity of visual approximation. We design two complementary tasks: Outline Prediction, which demands inferring global shapes from local components, and End-to-End Code Generation, which requires solving inverse geometric assembly problems. We conduct extensive evaluation experiments on advanced open-source and proprietary models, revealing an interesting insight: MLLMs tend to prioritize matching the target silhouette while neglecting geometric constraints, leading to distortions or deformations of the pieces.

Open-World Test-Time Training: Self-Training with Contrast Learning

Traditional test-time training (TTT) methods, while addressing domain shifts, often assume a consistent class set, limiting their applicability in real-world scenarios characterized by infinite variety. Open-World Test-Time Training (OWTTT) addresses the challenge of generalizing deep learning models to unknown target domain distributions, especially in the presence of strong Out-of-Distribution (OOD) data. Existing TTT methods often struggle to maintain performance when confronted with strong OOD data. In OWTTT, the focus has predominantly been on distinguishing between overall strong and weak OOD data. However, during the early stages of TTT, initial feature extraction is hampered by interference from strong OOD and corruptions, resulting in diminished contrast and premature classification of certain classes as strong OOD. To address this, we introduce Open World Dynamic Contrastive Learning (OWDCL), an innovative approach that utilizes contrastive learning to augment positive sample pairs. This strategy not only bolsters contrast in the early stages but also significantly enhances model robustness in subsequent stages. In comparison datasets, our OWDCL model has produced the most advanced performance.

Singular Value Penalization and Semantic Data Augmentation for Fully Test-Time Adaptation

Fully test-time adaptation (FTTA) adapts a model that is trained on a source domain to a target domain during the testing phase, where the two domains follow different distributions and source data is unavailable during the training phase. Existing methods usually adopt entropy minimization to reduce the uncertainty of target prediction results, and improve the FTTA performance accordingly. However, they fail to ensure the diversity in target prediction results. Recent domain adaptation study has shown that maximizing the sum of singular values of prediction results can simultaneously enhance their confidence (discriminability) and diversity. However, during the training phase, larger singular values usually take up a dominant position in loss maximization. This results in the model being more inclined to enhance discriminability for easily distinguishable classes, and the improvement in diversity is insufficiently effective. Furthermore, the adaptation and prediction in FTTA only use data from the current batch, which may lead to the risk of overfitting. To address the aforementioned issues, we propose maximizing the sum of singular values while minimizing their variance. This enables the model's focus toward the smaller singular values, enhancing discriminability between more challenging classes and effectively increasing the diversity of prediction results. Moreover, we incorporate data from the previous batch to realize semantic data augmentation for the current batch, reducing the risk of overfitting. Extensive experiments on benchmark datasets show our proposed approach outperforms some compared state-of-the-art FTTA methods.

Graph based Label Enhancement for Multi-instance Multi-label learning

Multi-instance multi-label (MIML) learning is widely applicated in numerous domains, such as the image classification where one image contains multiple instances correlated with multiple logic labels simultaneously. The related labels in existing MIML are all assumed as logical labels with equal significance. However, in practical applications in MIML, significance of each label for multiple instances per bag (such as an image) is significant different. Ignoring labeling significance will greatly lose the semantic information of the object, so that MIML is not applicable in complex scenes with a poor learning performance. To this end, this paper proposed a novel MIML framework based on graph label enhancement, namely GLEMIML, to improve the classification performance of MIML by leveraging label significance. GLEMIML first recognizes the correlations among instances by establishing the graph and then migrates the implicit information mined from the feature space to the label space via nonlinear mapping, thus recovering the label significance. Finally, GLEMIML is trained on the enhanced data through matching and interaction mechanisms. GLEMIML (AvgRank: 1.44) can effectively improve the performance of MIML by mining the label distribution mechanism and show better results than the SOTA method (AvgRank: 2.92) on multiple benchmark datasets.