Xueqi Ma

Despite excelling on multimodal benchmarks, vision-language models (VLMs) largely remain a black box. In this paper, we propose a novel interpretability framework to systematically analyze the internal mechanisms of VLMs, focusing on the functional roles of attention heads in multimodal reasoning. To this end, we introduce CogVision, a dataset that decomposes complex multimodal questions into step-by-step subquestions designed to simulate human reasoning through a chain-of-thought paradigm, with each subquestion associated with specific receptive or cognitive functions such as high-level visual reception and inference. Using a probing-based methodology, we identify attention heads that specialize in these functions and characterize them as functional heads. Our analysis across diverse VLM families reveals that these functional heads are universally sparse, vary in number and distribution across functions, and mediate interactions and hierarchical organization. Furthermore, intervention experiments demonstrate their critical role in multimodal reasoning: removing functional heads leads to performance degradation, while emphasizing them enhances accuracy. These findings provide new insights into the cognitive organization of VLMs and suggest promising directions for designing models with more human-aligned perceptual and reasoning abilities.

Text-to-image person retrieval (TIPR) aims to find images of the same identity that match a given text description. Current TIPR methods mainly focus on mining the association between images and texts, ignoring their potential complementarity. Besides, existing matching losses treat all positive pairs from the same identity equally, leading to noisy correspondences. In this paper, we propose CoRL: a cross-modal Collaborative Representation Learning framework designed to improve TIPR by effectively leveraging the complementarity between modalities. The text typically contains identity details with less noise, which helps distinguish visually similar pedestrians. This inspires us to integrate it into the corresponding image to emphasize identity-related and modality-shared visual information. However, corresponding text for each image is not always available, especially during inference. Accordingly, we introduce a Virtual-text Embedding Synthesizer that generates high-quality virtual-text features for cross-modal collaboration, eliminating the need for actual texts. We then design a Cross-Modal Collaboration learning process, incorporating a Cross-modal Relation Consistency loss to promote interaction and fusion between image and virtual-text features for mutual enhancement. Additionally, an Identity-bounded Matching loss is proposed to handle different types of image-text pairs distinctly, leading to more accurate cross-modal correspondences. Extensive experiments on multiple benchmarks demonstrate the superiority of CoRL over existing TIPR methods.

Graph masked autoencoders (GMAE) have emerged as a significant advancement in self-supervised pre-training for graph-structured data. Previous GMAE models primarily utilize a straightforward random masking strategy for nodes or edges during training. However, this strategy fails to consider the varying significance of different nodes within the graph structure. In this paper, we investigate the potential of leveraging the graph's structural composition as a fundamental and unique prior in the masked pre-training process. To this end, we introduce a novel structure-guided masking strategy (i.e., StructMAE), designed to refine the existing GMAE models. StructMAE involves two steps: 1) Structure-based Scoring: Each node is evaluated and assigned a score reflecting its structural significance. Two distinct types of scoring manners are proposed: predefined and learnable scoring. 2) Structure-guided Masking: With the obtained assessment scores, we develop an easy-to-hard masking strategy that gradually increases the structural awareness of the self-supervised reconstruction task. Specifically, the strategy begins with random masking and progresses to masking structure-informative nodes based on the assessment scores. This design gradually and effectively guides the model in learning graph structural information. Furthermore, extensive experiments consistently demonstrate that our StructMAE method outperforms existing state-of-the-art GMAE models in both unsupervised and transfer learning tasks. Codes are available at https://github.com/LiuChuang0059/StructMAE.

Graph Transformers (GTs) have demonstrated their advantages across a wide range of tasks. However, the self-attention mechanism in GTs overlooks the graph's inductive biases, particularly biases related to structure, which are crucial for the graph tasks. Although some methods utilize positional encoding and attention bias to model inductive biases, their effectiveness is still suboptimal analytically. Therefore, this paper presents Gradformer, a method innovatively integrating GT with the intrinsic inductive bias by applying an exponential decay mask to the attention matrix. Specifically, the values in the decay mask matrix diminish exponentially, correlating with the decreasing node proximities within the graph structure. This design enables Gradformer to retain its ability to capture information from distant nodes while focusing on the graph's local details. Furthermore, Gradformer introduces a learnable constraint into the decay mask, allowing different attention heads to learn distinct decay masks. Such an design diversifies the attention heads, enabling a more effective assimilation of diverse structural information within the graph. Extensive experiments on various benchmarks demonstrate that Gradformer consistently outperforms the Graph Neural Network and GT baseline models in various graph classification and regression tasks. Additionally, Gradformer has proven to be an effective method for training deep GT models, maintaining or even enhancing accuracy compared to shallow models as the network deepens, in contrast to the significant accuracy drop observed in other GT models.Codes are available at \url{https://github.com/LiuChuang0059/Gradformer}.

Graph Transformers (GTs) have achieved impressive results on various graph-related tasks. However, the huge computational cost of GTs hinders their deployment and application, especially in resource-constrained environments. Therefore, in this paper, we explore the feasibility of sparsifying GTs, a significant yet under-explored topic. We first discuss the redundancy of GTs based on the characteristics of existing GT models, and then propose a comprehensive \textbf{G}raph \textbf{T}ransformer \textbf{SP}arsification (GTSP) framework that helps to reduce the computational complexity of GTs from four dimensions: the input graph data, attention heads, model layers, and model weights. Specifically, GTSP designs differentiable masks for each individual compressible component, enabling effective end-to-end pruning. We examine our GTSP through extensive experiments on prominent GTs, including GraphTrans, Graphormer, and GraphGPS. The experimental results substantiate that GTSP effectively cuts computational costs, accompanied by only marginal decreases in accuracy or, in some cases, even improvements. For instance, GTSP yields a reduction of 30\% in Floating Point Operations while contributing to a 1.8\% increase in Area Under the Curve accuracy on OGBG-HIV dataset. Furthermore, we provide several insights on the characteristics of attention heads and the behavior of attention mechanisms, all of which have immense potential to inspire future research endeavors in this domain.