In this paper, we address the challenging task of multimodal mathematical reasoning by incorporating the ability of "slow thinking" into multimodal large language models (MLLMs). Our core idea is that different levels of reasoning abilities can be combined dynamically to tackle questions with different complexity. To this end, we propose a paradigm of Self-structured Chain of Thought (SCoT), which is composed of minimal semantic atomic steps. Different from existing methods that rely on structured templates or free-form paradigms, our method can not only generate cognitive CoT structures for various complex tasks but also mitigates the phenomenon of overthinking. To introduce structured reasoning capabilities into visual understanding models, we further design a novel AtomThink framework with four key modules, including (i) a data engine to generate high-quality multimodal reasoning paths; (ii) a supervised fine-tuning process with serialized inference data; (iii) a policy-guided multi-turn inference method; and (iv) an atomic capability metric to evaluate the single step utilization rate. We conduct extensive experiments to show that the proposed AtomThink significantly improves the performance of baseline MLLMs, achieving more than 10\% average accuracy gains on MathVista and MathVerse. Compared to state-of-the-art structured CoT approaches, our method not only achieves higher accuracy but also improves data utilization by 5 times and boosts inference efficiency by 85.3\%. Our code is now public available in https://github.com/Quinn777/AtomThink.

In this paper, we address the challenging task of multimodal mathematical reasoning by incorporating the ability of ``slow thinking" into multimodal large language models (MLLMs). Contrary to existing methods that rely on direct or fast thinking, our key idea is to construct long chains of thought (CoT) consisting of atomic actions in a step-by-step manner, guiding MLLMs to perform complex reasoning. To this end, we design a novel AtomThink framework composed of three key modules: (i) a CoT annotation engine that automatically generates high-quality CoT annotations to address the lack of high-quality visual mathematical data; (ii) an atomic step fine-tuning strategy that jointly optimizes an MLLM and a policy reward model (PRM) for step-wise reasoning; and (iii) four different search strategies that can be applied with the PRM to complete reasoning. Additionally, we propose AtomMATH, a large-scale multimodal dataset of long CoTs, and an atomic capability evaluation metric for mathematical tasks. Extensive experimental results show that the proposed AtomThink significantly improves the performance of baseline MLLMs, achieving approximately 50\% relative accuracy gains on MathVista and 120\% on MathVerse. To support the advancement of multimodal slow-thinking models, we will make our code and dataset publicly available on https://github.com/Quinn777/AtomThink.

GPT-4o, an omni-modal model that enables vocal conversations with diverse emotions and tones, marks a milestone for omni-modal foundation models. However, empowering Large Language Models to perceive and generate images, texts, and speeches end-to-end with publicly available data remains challenging in the open-source community. Existing vision-language models rely on external tools for the speech processing, while speech-language models still suffer from limited or even without vision-understanding abilities. To address this gap, we propose EMOVA (EMotionally Omni-present Voice Assistant), to enable Large Language Models with end-to-end speech capabilities while maintaining the leading vision-language performance. With a semantic-acoustic disentangled speech tokenizer, we notice surprisingly that omni-modal alignment can further enhance vision-language and speech abilities compared with the corresponding bi-modal aligned counterparts. Moreover, a lightweight style module is proposed for flexible speech style controls (e.g., emotions and pitches). For the first time, EMOVA achieves state-of-the-art performance on both the vision-language and speech benchmarks, and meanwhile, supporting omni-modal spoken dialogue with vivid emotions.

As the capabilities of large language models (LLMs) have expanded dramatically, aligning these models with human values presents a significant challenge. Traditional alignment strategies rely heavily on human intervention, such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), or on the self-alignment capacities of LLMs, which usually require a strong LLM's emergent ability to improve its original bad answer. To address these challenges, we propose a novel self-alignment method that utilizes a Chain of Thought (CoT) approach, termed AlignCoT. This method encompasses stages of Question Analysis, Answer Guidance, and Safe Answer production. It is designed to enable LLMs to generate high-quality, safe responses throughout various stages of their development. Furthermore, we introduce the Mixture of insighTful Experts (MoTE) architecture, which applies mixture of experts to enhance each component of the AlignCoT process, markedly increasing alignment efficiency. The MoTE approach not only outperforms existing methods in aligning LLMs with human values but also highlights the benefits of using self-generated data, revealing the dual benefits of improved alignment and training efficiency.

Vision-language pre-trained models have achieved impressive performance on various downstream tasks. However, their large model sizes hinder their utilization on platforms with limited computational resources. We find that directly using smaller pre-trained models and applying magnitude-based pruning on CLIP models leads to inflexibility and inferior performance. Recent efforts for VLP compression either adopt uni-modal compression metrics resulting in limited performance or involve costly mask-search processes with learnable masks. In this paper, we first propose the Module-wise Pruning Error (MoPE) metric, accurately assessing CLIP module importance by performance decline on cross-modal tasks. Using the MoPE metric, we introduce a unified pruning framework applicable to both pre-training and task-specific fine-tuning compression stages. For pre-training, MoPE-CLIP effectively leverages knowledge from the teacher model, significantly reducing pre-training costs while maintaining strong zero-shot capabilities. For fine-tuning, consecutive pruning from width to depth yields highly competitive task-specific models. Extensive experiments in two stages demonstrate the effectiveness of the MoPE metric, and MoPE-CLIP outperforms previous state-of-the-art VLP compression methods.

Large Language Models (LLMs) have exhibited remarkable success in long-form context comprehension tasks. However, their capacity to generate long contents, such as reports and articles, remains insufficiently explored. Current benchmarks do not adequately assess LLMs' ability to produce informative and comprehensive content, necessitating a more rigorous evaluation approach. In this study, we introduce \textsc{ProxyQA}, a framework for evaluating long-form text generation, comprising in-depth human-curated \textit{meta-questions} spanning various domains. Each meta-question contains corresponding \textit{proxy-questions} with annotated answers. LLMs are prompted to generate extensive content in response to these meta-questions. Utilizing an evaluator and incorporating generated content as background context, \textsc{ProxyQA} evaluates the quality of generated content based on the evaluator's performance in answering the \textit{proxy-questions}. We examine multiple LLMs, emphasizing \textsc{ProxyQA}'s demanding nature as a high-quality assessment tool. Human evaluation demonstrates that evaluating through \textit{proxy-questions} is a highly self-consistent and human-criteria-correlated validation method. The dataset and leaderboard will be available at \url{https://github.com/Namco0816/ProxyQA}.

Masked Autoencoder (MAE) is a prevailing self-supervised learning method that achieves promising results in model pre-training. However, when the various downstream tasks have data distributions different from the pre-training data, the semantically irrelevant pre-training information might result in negative transfer, impeding MAE's scalability. To address this issue, we propose a novel MAEbased pre-training paradigm, Mixture of Cluster-conditional Experts (MoCE), which can be trained once but provides customized pre-training models for diverse downstream tasks. Thus, each downstream task can be allocated to its customized model pretrained with data most similar to the downstream data. Experiments on a collection of 11 downstream tasks show that MoCE outperforms the vanilla MAE by 2.45% on average. Self-supervised learning (SSL), which learns effective transferable representations without human annotations, has become a prevailing model pre-training paradigm (He et al., 2020; Chen et al., 2021a; Bao et al., 2022). Currently, the most prevalent SSL method is the Masked Autoencoder (MAE) (He et al., 2022), which constructs supervision signals from raw image data by masking random input patches and then reconstructing the missing pixels. This simple strategy has proved efficient in the training of large-scale models. For example, ViT (Dosovitskiy et al., 2021) shows impressive performance on popular benchmarks such as the ImageNet However, does MAE really scale well for various downstream tasks (Deng et al., 2009; Lin et al., 2014; Zhou et al., 2019; Han et al., 2021; Li et al., 2022a)? Preliminary studies (in Section 3.1) show that the MAE indeed suffers from negative transfer (Liu et al., 2022) when transferring to downstream tasks with very different semantics. Figure 1(a) shows that on 9 of 11 downstream tasks, an MAE pre-trained on the full ImageNet data is outperformed by the one that is pre-trained on only the semantically relevant data subsets.

Diffusion models have gained prominence in generating data for perception tasks such as image classification and object detection. However, the potential in generating high-quality tracking sequences, a crucial aspect in the field of video perception, has not been fully investigated. To address this gap, we propose TrackDiffusion, a novel architecture designed to generate continuous video sequences from the tracklets. TrackDiffusion represents a significant departure from the traditional layout-to-image (L2I) generation and copy-paste synthesis focusing on static image elements like bounding boxes by empowering image diffusion models to encompass dynamic and continuous tracking trajectories, thereby capturing complex motion nuances and ensuring instance consistency among video frames. For the first time, we demonstrate that the generated video sequences can be utilized for training multi-object tracking (MOT) systems, leading to significant improvement in tracker performance. Experimental results show that our model significantly enhances instance consistency in generated video sequences, leading to improved perceptual metrics. Our approach achieves an improvement of 8.7 in TrackAP and 11.8 in TrackAP$_{50}$ on the YTVIS dataset, underscoring its potential to redefine the standards of video data generation for MOT tasks and beyond.

Masked Autoencoder (MAE) has demonstrated superior performance on various vision tasks via randomly masking image patches and reconstruction. However, effective data augmentation strategies for MAE still remain open questions, different from those in contrastive learning that serve as the most important part. This paper studies the prevailing mixing augmentation for MAE. We first demonstrate that naive mixing will in contrast degenerate model performance due to the increase of mutual information (MI). To address, we propose homologous recognition, an auxiliary pretext task, not only to alleviate the MI increasement by explicitly requiring each patch to recognize homologous patches, but also to perform object-aware self-supervised pre-training for better downstream dense perception performance. With extensive experiments, we demonstrate that our proposed Mixed Autoencoder (MixedAE) achieves the state-of-the-art transfer results among masked image modeling (MIM) augmentations on different downstream tasks with significant efficiency. Specifically, our MixedAE outperforms MAE by +0.3% accuracy, +1.7 mIoU and +0.9 AP on ImageNet-1K, ADE20K and COCO respectively with a standard ViT-Base. Moreover, MixedAE surpasses iBOT, a strong MIM method combined with instance discrimination, while accelerating training by 2x. To our best knowledge, this is the very first work to consider mixing for MIM from the perspective of pretext task design. Code will be made available.

Contrastive learning, especially self-supervised contrastive learning (SSCL), has achieved great success in extracting powerful features from unlabeled data. In this work, we contribute to the theoretical understanding of SSCL and uncover its connection to the classic data visualization method, stochastic neighbor embedding (SNE) (Hinton & Roweis, 2002), whose goal is to preserve pairwise distances. From the perspective of preserving neighboring information, SSCL can be viewed as a special case of SNE with the input space pairwise similarities specified by data augmentation. The established correspondence facilitates deeper theoretical understanding of learned features of SSCL, as well as methodological guidelines for practical improvement. Specifically, through the lens of SNE, we provide novel analysis on domain-agnostic augmentations, implicit bias and robustness of learned features. To illustrate the practical advantage, we demonstrate that the modifications from SNE to t-SNE (Van der Maaten & Hinton, 2008) can also be adopted in the SSCL setting, achieving significant improvement in both in-distribution and out-of-distribution generalization. In contrast to supervised learning, SSCL learns the representation through a large number of unlabeled data and artificially defined self-supervision signals, i.e., regarding the augmented views of a data sample as positive pairs and randomly sampled data as negative pairs. By enforcing the features of positive pairs to align and those of negative pairs to be distant, SSCL produces discriminative features with state-of-the-art performance for various downstream tasks. Despite the empirical success, the theoretical understanding is under-explored as to how the learned features depend on the data and augmentation, how different components in SSCL work and what are the implicit biases when there exist multiple empirical loss minimizers. For instance, SSCL methods are widely adopted for pretraining, whose feature mappings are to be utilized for various downstream tasks which are usually out-of-distribution (OOD). The distribution shift poses great challenges for the feature learning process with extra requirement for robustness and OOD generalization (Arjovsky et al., 2019; Krueger et al., 2021; Bai et al., 2021; He et al., 2020b; Zhao et al., 2023; Dong et al., 2022), which demands deeper understanding of the SSCL methods. The goal of SSCL is to learn the feature representations from data. For this problem, one classic method is SNE (Hinton et al., 2006) and its various extensions.