Tao Chen

Jianjian Cao, Peng Ye, Shengze Li et al.

Vision-Language Transformers (VLTs) have shown great success recently, but are meanwhile accompanied by heavy computation costs, where a major reason can be attributed to the large number of visual and language tokens. Existing token pruning research for compressing VLTs mainly follows a single-modality-based scheme yet ignores the critical role of aligning different modalities for guiding the token pruning process, causing the important tokens for one modality to be falsely pruned in another modality branch. Meanwhile, existing VLT pruning works also lack the flexibility to dynamically compress each layer based on different input samples. To this end, we propose a novel framework named Multimodal Alignment-Guided Dynamic Token Pruning (MADTP) for accelerating various VLTs. Specifically, we first introduce a well-designed Multi-modality Alignment Guidance (MAG) module that can align features of the same semantic concept from different modalities, to ensure the pruned tokens are less important for all modalities. We further design a novel Dynamic Token Pruning (DTP) module, which can adaptively adjust the token compression ratio in each layer based on different input instances. Extensive experiments on various benchmarks demonstrate that MADTP significantly reduces the computational complexity of kinds of multimodal models while preserving competitive performance. Notably, when applied to the BLIP model in the NLVR2 dataset, MADTP can reduce the GFLOPs by 80% with less than 4% performance degradation.

Peng Ye, Chenyu Huang, Mingzhu Shen et al.

This work targets to merge various Vision Transformers (ViTs) trained on different tasks (i.e., datasets with different object categories) or domains (i.e., datasets with the same categories but different environments) into one unified model, yielding still good performance on each task or domain. Previous model merging works focus on either CNNs or NLP models, leaving the ViTs merging research untouched. To fill this gap, we first explore and find that existing model merging methods cannot well handle the merging of the whole ViT models and still have improvement space. To enable the merging of the whole ViT, we propose a simple-but-effective gating network that can both merge all kinds of layers (e.g., Embedding, Norm, Attention, and MLP) and select the suitable classifier. Specifically, the gating network is trained by unlabeled datasets from all the tasks (domains), and predicts the probability of which task (domain) the input belongs to for merging the models during inference. To further boost the performance of the merged model, especially when the difficulty of merging tasks increases, we design a novel metric of model weight similarity, and utilize it to realize controllable and combined weight merging. Comprehensive experiments on kinds of newly established benchmarks, validate the superiority of the proposed ViT merging framework for different tasks and domains. Our method can even merge beyond 10 ViT models from different vision tasks with a negligible effect on the performance of each task.

Chongjun Tu, Lin Zhang, Pengtao Chen et al.

Multimodal Large Language Models (MLLMs) have shown remarkable capabilities in video content understanding but still struggle with fine-grained motion comprehension. To comprehensively assess the motion understanding ability of existing MLLMs, we introduce FAVOR-Bench, comprising 1,776 videos with structured manual annotations of various motions. Our benchmark includes both close-ended and open-ended tasks. For close-ended evaluation, we carefully design 8,184 multiple-choice question-answer pairs spanning six distinct sub-tasks. For open-ended evaluation, we develop both a novel cost-efficient LLM-free and a GPT-assisted caption assessment method, where the former can enhance benchmarking interpretability and reproducibility. Comprehensive experiments with 21 state-of-the-art MLLMs reveal significant limitations in their ability to comprehend and describe detailed temporal dynamics in video motions. To alleviate this limitation, we further build FAVOR-Train, a dataset consisting of 17,152 videos with fine-grained motion annotations. The results of finetuning Qwen2.5-VL on FAVOR-Train yield consistent improvements on motion-related tasks of TVBench, MotionBench and our FAVOR-Bench. Comprehensive assessment results demonstrate that the proposed FAVOR-Bench and FAVOR-Train provide valuable tools to the community for developing more powerful video understanding models. Project page: \href{https://favor-bench.github.io/}{https://favor-bench.github.io/}.

Minglei Li, Peng Ye, Yongqi Huang et al.

Parameter-efficient fine-tuning (PEFT) has become increasingly important as foundation models continue to grow in both popularity and size. Adapter has been particularly well-received due to their potential for parameter reduction and adaptability across diverse tasks. However, striking a balance between high efficiency and robust generalization across tasks remains a challenge for adapter-based methods. We analyze existing methods and find that: 1) parameter sharing is the key to reducing redundancy; 2) more tunable parameters, dynamic allocation, and block-specific design are keys to improving performance. Unfortunately, no previous work considers all these factors. Inspired by this insight, we introduce a novel framework named Adapter-X. First, a Sharing Mixture of Adapters (SMoA) module is proposed to fulfill token-level dynamic allocation, increased tunable parameters, and inter-block sharing at the same time. Second, some block-specific designs like Prompt Generator (PG) are introduced to further enhance the ability of adaptation. Extensive experiments across 2D image and 3D point cloud modalities demonstrate that Adapter-X represents a significant milestone as it is the first to outperform full fine-tuning in both 2D image and 3D point cloud modalities with significantly fewer parameters, i.e., only 0.20% and 1.88% of original trainable parameters for 2D and 3D classification tasks. Our code will be publicly available.