Abstract:The rapid development of Artificial Intelligence (AI) has revolutionized numerous fields, with large language models (LLMs) and computer vision (CV) systems driving advancements in natural language understanding and visual processing, respectively. The convergence of these technologies has catalyzed the rise of multimodal AI, enabling richer, cross-modal understanding that spans text, vision, audio, and video modalities. Multimodal large language models (MLLMs), in particular, have emerged as a powerful framework, demonstrating impressive capabilities in tasks like image-text generation, visual question answering, and cross-modal retrieval. Despite these advancements, the complexity and scale of MLLMs introduce significant challenges in interpretability and explainability, essential for establishing transparency, trustworthiness, and reliability in high-stakes applications. This paper provides a comprehensive survey on the interpretability and explainability of MLLMs, proposing a novel framework that categorizes existing research across three perspectives: (I) Data, (II) Model, (III) Training \& Inference. We systematically analyze interpretability from token-level to embedding-level representations, assess approaches related to both architecture analysis and design, and explore training and inference strategies that enhance transparency. By comparing various methodologies, we identify their strengths and limitations and propose future research directions to address unresolved challenges in multimodal explainability. This survey offers a foundational resource for advancing interpretability and transparency in MLLMs, guiding researchers and practitioners toward developing more accountable and robust multimodal AI systems.
Abstract:In recent years, multimodal large language models (MLLMs) have significantly advanced, integrating more modalities into diverse applications. However, the lack of explainability remains a major barrier to their use in scenarios requiring decision transparency. Current neuron-level explanation paradigms mainly focus on knowledge localization or language- and domain-specific analyses, leaving the exploration of multimodality largely unaddressed. To tackle these challenges, we propose MINER, a transferable framework for mining modality-specific neurons (MSNs) in MLLMs, which comprises four stages: (1) modality separation, (2) importance score calculation, (3) importance score aggregation, (4) modality-specific neuron selection. Extensive experiments across six benchmarks and two representative MLLMs show that (I) deactivating ONLY 2% of MSNs significantly reduces MLLMs performance (0.56 to 0.24 for Qwen2-VL, 0.69 to 0.31 for Qwen2-Audio), (II) different modalities mainly converge in the lower layers, (III) MSNs influence how key information from various modalities converges to the last token, (IV) two intriguing phenomena worth further investigation, i.e., semantic probing and semantic telomeres. The source code is available at this URL.
Abstract:As the field of Multimodal Large Language Models (MLLMs) continues to evolve, their potential to revolutionize artificial intelligence is particularly promising, especially in addressing mathematical reasoning tasks. Current mathematical benchmarks predominantly focus on evaluating MLLMs' problem-solving ability, yet there is a crucial gap in addressing more complex scenarios such as error detection, for enhancing reasoning capability in complicated settings. To fill this gap, we formally formulate the new task: multimodal error detection, and introduce ErrorRadar, the first benchmark designed to assess MLLMs' capabilities in such a task. ErrorRadar evaluates two sub-tasks: error step identification and error categorization, providing a comprehensive framework for evaluating MLLMs' complex mathematical reasoning ability. It consists of 2,500 high-quality multimodal K-12 mathematical problems, collected from real-world student interactions in an educational organization, with rigorous annotation and rich metadata such as problem type and error category. Through extensive experiments, we evaluated both open-source and closed-source representative MLLMs, benchmarking their performance against educational expert evaluators. Results indicate significant challenges still remain, as GPT-4o with best performance is still around 10% behind human evaluation. The dataset will be available upon acceptance.
Abstract:Projecting visual features into word embedding space has become a significant fusion strategy adopted by Multimodal Large Language Models (MLLMs). However, its internal mechanisms have yet to be explored. Inspired by multilingual research, we identify domain-specific neurons in multimodal large language models. Specifically, we investigate the distribution of domain-specific neurons and the mechanism of how MLLMs process features from diverse domains. Furthermore, we propose a three-stage framework for language model modules in MLLMs when handling projected image features, and verify this hypothesis using logit lens. Extensive experiments indicate that while current MLLMs exhibit Visual Question Answering (VQA) capability, they may not fully utilize domain-specific information. Manipulating domain-specific neurons properly will result in a 10\% change of accuracy at most, shedding light on the development of cross-domain, all-encompassing MLLMs in the future. Our code will be released upon paper notification.
Abstract:Catastrophic forgetting in neural networks during incremental learning remains a challenging problem. Previous research investigated catastrophic forgetting in fully connected networks, with some earlier work exploring activation functions and learning algorithms. Applications of neural networks have been extended to include similarity learning. It is of significant interest to understand how similarity learning loss functions would be affected by catastrophic forgetting. Our research investigates catastrophic forgetting for four well-known similarity-based loss functions during incremental class learning. The loss functions are angular, contrastive, centre, and triplet loss. Our results show that the rate of catastrophic forgetting is different across loss functions on multiple datasets. The angular loss was least affected, followed by contrastive, triplet loss, and centre loss with good mining techniques. We implemented three existing incremental learning techniques, iCaRL, EWC, and EBLL. We further proposed our novel technique using VAEs to generate representation as exemplars that are passed through intermediate layers of the network. Our method outperformed the three existing techniques. We have shown that we do not require stored images as exemplars for incremental learning with similarity learning. The generated representations can help preserve regions of the embedding space used by prior knowledge so that new knowledge will not ``overwrite'' prior knowledge.
Abstract:This paper tackles face recognition in videos employing metric learning methods and similarity ranking models. The paper compares the use of the Siamese network with contrastive loss and Triplet Network with triplet loss implementing the following architectures: Google/Inception architecture, 3D Convolutional Network (C3D), and a 2-D Long short-term memory (LSTM) Recurrent Neural Network. We make use of still images and sequences from videos for training the networks and compare the performances implementing the above architectures. The dataset used was the YouTube Face Database designed for investigating the problem of face recognition in videos. The contribution of this paper is two-fold: to begin, the experiments have established 3-D Convolutional networks and 2-D LSTMs with the contrastive loss on image sequences do not outperform Google/Inception architecture with contrastive loss in top $n$ rank face retrievals with still images. However, the 3-D Convolution networks and 2-D LSTM with triplet Loss outperform the Google/Inception with triplet loss in top $n$ rank face retrievals on the dataset; second, a Support Vector Machine (SVM) was used in conjunction with the CNNs' learned feature representations for facial identification. The results show that feature representation learned with triplet loss is significantly better for n-shot facial identification compared to contrastive loss. The most useful feature representations for facial identification are from the 2-D LSTM with triplet loss. The experiments show that learning spatio-temporal features from video sequences is beneficial for facial recognition in videos.