As the capabilities of Multimodal Large Language Models (MLLMs) improve, the need for higher-order evaluation of them is increasing. However, there is a lack of work evaluating MLLM for higher-order perception and understanding of Chinese visual content. To address this, we introduce the CII-Bench, which aims to assess MLLMs’ such capabilities for Chinese images. To ensure the authenticity of the Chinese context, images in CII-Bench are sourced from the Chinese Internet and manually reviewed, with corresponding answers also manually crafted. Additionally, CII-Bench incorporates images that represent Chinese traditional culture, such as famous Chinese traditional paintings, which can deeply reflect the model’s understanding of Chinese traditional culture. Through experiments on multiple MLLMs using CII-Bench, significant findings emerged. There is a large gap between MLLMs and humans in performance. The highest MLLM accuracy is 64.4%, while the human average is 78.2% and the peak is 81.0%. MLLMs perform poorly on traditional culture images, indicating limitations in understanding high-level semantics and lacking a deep knowledge base of Chinese traditional culture. Moreover, most models have higher accuracy when image emotion hints are added to the prompts. We believe CII-Bench will help MLLMs better understand Chinese semantics and specific images, and move forward the development of expert artificial general intelligence (AGI). Our project is publicly available at https://cii-bench.github.io.

Fully test-time adaptation (FTTA) is to adapt a model trained on a source domain to a target domain during the testing phase. Traditional methods like entropy minimization primarily focus on reducing uncertainty in output predictions, yet often overlook the diversity in target prediction results, which is critical for unbalanced classes in complex datasets. To address this, our study introduces a new method named Variance Balancing and Semantic Augmentation (VBSA). VBSA begins by maximizing the sum of singular values in predictions, coupled with a novel variance penalization strategy. This strategy not only focuses the model on unbalanced classes but also mitigates the overfitting risk associated with singular value maximization, thereby ensuring a balanced emphasis across various classes and enhancing the diversity of prediction results. Furthermore, VBSA incorporates semantic data augmentation using data from previous batches, offering semantic-level augmentation for all classes, with particular benefits for unbalanced ones. Extensive experiments demonstrate that our VBSA method has produced the most advanced performance.

Domain adaptation (DA) is crucial for transferring models trained in one domain to perform well in a different, often unseen domain. Traditional methods, including unsupervised domain adaptation (UDA) and source-free domain adaptation (SFDA), have made significant progress. However, most existing DA methods rely heavily on Batch Normalization (BN) layers, which are not optimal in source-free settings, where the source domain is unavailable for comparison. In this study, we propose a novel method, ESBN, which addresses the challenge of domain shift by adjusting the placement of normalization layers and replacing BN with Batch-free Normalization (BFN). Unlike BN, BFN is less dependent on batch statistics and provides more robust feature representations through instance-specific statistics. We systematically investigate the effects of different BN layer placements across various network configurations and demonstrate that selective replacement with BFN improves generalization performance. Extensive experiments on multiple domain adaptation benchmarks show that our approach outperforms state-of-the-art methods, particularly in challenging scenarios such as Open-Partial Domain Adaptation (OPDA).

In computer vision, deep learning models excel in tasks with data adhering to the independently and identically distributed principle but struggle with distribution shifts between source and target domains. Domain Generalization (DG) aims to train models on source data for effective performance on unseen domains. This paper addresses two pivotal challenges in DG: the tendency of models to overly focus on transferability at the cost of distinguishability, and the common oversight of imbalanced classes in long-tailed datasets. We introduce an innovative approach that employs singular value maximization and suppression techniques. This method strikes a balance between transferability and distinguishability, while also emphasizing indistinct and unbalanced classes in long-tailed datasets. Our approach utilizes distinct maximum singular value suppression techniques at both the feature and output levels. At the feature level, our method balances transferability and distinguishability, offering robustness superior to previous studies. At the batch level, maximum singular value suppression is tailored to enhance focus on imbalanced and indistinct classes. Extensive experimental results across multiple datasets demonstrate our approach’s superior performance over existing DG techniques.

Despite progress, deep neural networks still suffer performance declines under distribution shifts between training and test domains, leading to a substantial decrease in Quality of Experience (QoE) for applications. Existing test-time adaptation (TTA) methods are challenged by dynamic, multiple test distributions within batches. We observe that feature distributions across different domains inherently cluster into distinct groups with varying means and variances. This divergence reveals a critical limitation of previous global normalization strategies in TTA, which inevitably distort the original data characteristics. Based on this insight, we propose Feature-based Instance Neighbor Discovery (FIND), which comprises three key components: Layer-Wise Feature Disentanglement (LFD), Feature-Aware Batch Normalization (FABN) and Selective FABN (S-FABN). LFD stably captures features with similar distributions at each layer by constructing graph structures; while FABN optimally combines source statistics with test-time distribution-specific statistics for robust feature representation. Finally, S-FABN determines which layers require feature partitioning and which can remain unified, thus enhancing the efficiency of inference. Extensive experiments demonstrate that FIND significantly outperforms existing methods, achieving up to approximately 30\% accuracy improvement in dynamic scenarios while maintaining computational efficiency. The source code is available at https://github.com/Peanut-255/FIND.

Recent advancements in large language models (LLMs) underscore the need for more comprehensive evaluation methods to accurately assess their reasoning capabilities. Existing benchmarks are often domain-specific and thus cannot fully capture an LLM’s general reasoning potential. To address this limitation, we introduce the **Knowledge Orthogonal Reasoning Gymnasium (KORGym)**, a dynamic evaluation platform inspired by KOR-Bench and Gymnasium. KORGym offers over fifty games in either textual or visual formats and supports interactive, multi-turn assessments with reinforcement learning scenarios. Using KORGym, we conduct extensive experiments on 19 LLMs and 8 VLMs, revealing consistent reasoning patterns within model families and demonstrating the superior performance of closed-source models. Further analysis examines the effects of modality, reasoning strategies, reinforcement learning techniques, and response length on model performance. We expect KORGym to become a valuable resource for advancing LLM reasoning research and developing evaluation methodologies suited to complex, interactive environments.