Di Hu

Xiaokang Peng, Yake Wei, Andong Deng et al.

Multimodal learning helps to comprehensively understand the world, by integrating different senses. Accordingly, multiple input modalities are expected to boost model performance, but we actually find that they are not fully exploited even when the multimodal model outperforms its uni-modal counterpart. Specifically, in this paper we point out that existing multimodal discriminative models, in which uniform objective is designed for all modalities, could remain under-optimized uni-modal representations, caused by another dominated modality in some scenarios, e.g., sound in blowing wind event, vision in drawing picture event, etc. To alleviate this optimization imbalance, we propose on-the-fly gradient modulation to adaptively control the optimization of each modality, via monitoring the discrepancy of their contribution towards the learning objective. Further, an extra Gaussian noise that changes dynamically is introduced to avoid possible generalization drop caused by gradient modulation. As a result, we achieve considerable improvement over common fusion methods on different multimodal tasks, and this simple strategy can also boost existing multimodal methods, which illustrates its efficacy and versatility. The source code is available at \url{https://github.com/GeWu-Lab/OGM-GE_CVPR2022}.

Wenke Xia, Xingjian Li, Andong Deng et al.

Cross-modal distillation has been widely used to transfer knowledge across different modalities, enriching the representation of the target unimodal one. Recent studies highly relate the temporal synchronization between vision and sound to the semantic consistency for cross-modal distillation. However, such semantic consistency from the synchronization is hard to guarantee in unconstrained videos, due to the irrelevant modality noise and differentiated semantic correlation. To this end, we first propose a \textit{Modality Noise Filter} (MNF) module to erase the irrelevant noise in teacher modality with cross-modal context. After this purification, we then design a \textit{Contrastive Semantic Calibration} (CSC) module to adaptively distill useful knowledge for target modality, by referring to the differentiated sample-wise semantic correlation in a contrastive fashion. Extensive experiments show that our method could bring a performance boost compared with other distillation methods in both visual action recognition and video retrieval task. We also extend to the audio tagging task to prove the generalization of our method. The source code is available at \href{https://github.com/GeWu-Lab/cross-modal-distillation}{https://github.com/GeWu-Lab/cross-modal-distillation}.

Ziyun Li, Jona Otholt, Ben Dai et al.

Novel class discovery (NCD) aims to infer novel categories in an unlabeled dataset by leveraging prior knowledge of a labeled set comprising disjoint but related classes. Given that most existing literature focuses primarily on utilizing supervised knowledge from a labeled set at the methodology level, this paper considers the question: Is supervised knowledge always helpful at different levels of semantic relevance? To proceed, we first establish a novel metric, so-called transfer flow, to measure the semantic similarity between labeled/unlabeled datasets. To show the validity of the proposed metric, we build up a large-scale benchmark with various degrees of semantic similarities between labeled/unlabeled datasets on ImageNet by leveraging its hierarchical class structure. The results based on the proposed benchmark show that the proposed transfer flow is in line with the hierarchical class structure; and that NCD performance is consistent with the semantic similarities (measured by the proposed metric). Next, by using the proposed transfer flow, we conduct various empirical experiments with different levels of semantic similarity, yielding that supervised knowledge may hurt NCD performance. Specifically, using supervised information from a low-similarity labeled set may lead to a suboptimal result as compared to using pure self-supervised knowledge. These results reveal the inadequacy of the existing NCD literature which usually assumes that supervised knowledge is beneficial. Finally, we develop a pseudo-version of the transfer flow as a practical reference to decide if supervised knowledge should be used in NCD. Its effectiveness is supported by our empirical studies, which show that the pseudo transfer flow (with or without supervised knowledge) is consistent with the corresponding accuracy based on various datasets. Code is released at https://github.com/J-L-O/SK-Hurt-NCD

Jie Chen, Zhipeng Chen, Jiapeng Wang et al.

Continual pre-training (CPT) has been an important approach for adapting language models to specific domains or tasks. To make the CPT approach more traceable, this paper presents a technical report for continually pre-training Llama-3 (8B), which significantly enhances the Chinese language ability and scientific reasoning ability of the backbone model. To enhance the new abilities while retaining the original abilities, we design specific data mixture and curriculum strategies by utilizing existing datasets and synthesizing high-quality datasets. Specifically, we synthesize multidisciplinary scientific question and answer (QA) pairs based on related web pages, and subsequently incorporate these synthetic data to improve the scientific reasoning ability of Llama-3. We refer to the model after CPT as Llama-3-SynE (Synthetic data Enhanced Llama-3). We also present the tuning experiments with a relatively small model -- TinyLlama, and employ the derived findings to train the backbone model. Extensive experiments on a number of evaluation benchmarks show that our approach can largely improve the performance of the backbone models, including both the general abilities (+8.81 on C-Eval and +6.31 on CMMLU) and the scientific reasoning abilities (+12.00 on MATH and +4.13 on SciEval), without hurting the original capacities. Our model, data, and codes are available at https://github.com/RUC-GSAI/Llama-3-SynE.

Guangyao Li, Yake Wei, Yapeng Tian et al.

In this paper, we focus on the Audio-Visual Question Answering (AVQA) task, which aims to answer questions regarding different visual objects, sounds, and their associations in videos. The problem requires comprehensive multimodal understanding and spatio-temporal reasoning over audio-visual scenes. To benchmark this task and facilitate our study, we introduce a large-scale MUSIC-AVQA dataset, which contains more than 45K question-answer pairs covering 33 different question templates spanning over different modalities and question types. We develop several baselines and introduce a spatio-temporal grounded audio-visual network for the AVQA problem. Our results demonstrate that AVQA benefits from multisensory perception and our model outperforms recent A-, V-, and AVQA approaches. We believe that our built dataset has the potential to serve as testbed for evaluating and promoting progress in audio-visual scene understanding and spatio-temporal reasoning. Code and dataset: http://gewu-lab.github.io/MUSIC-AVQA/

Wenke Xia, Dong Wang, Xincheng Pang et al.

Generalizable articulated object manipulation is essential for home-assistant robots. Recent efforts focus on imitation learning from demonstrations or reinforcement learning in simulation, however, due to the prohibitive costs of real-world data collection and precise object simulation, it still remains challenging for these works to achieve broad adaptability across diverse articulated objects. Recently, many works have tried to utilize the strong in-context learning ability of Large Language Models (LLMs) to achieve generalizable robotic manipulation, but most of these researches focus on high-level task planning, sidelining low-level robotic control. In this work, building on the idea that the kinematic structure of the object determines how we can manipulate it, we propose a kinematic-aware prompting framework that prompts LLMs with kinematic knowledge of objects to generate low-level motion trajectory waypoints, supporting various object manipulation. To effectively prompt LLMs with the kinematic structure of different objects, we design a unified kinematic knowledge parser, which represents various articulated objects as a unified textual description containing kinematic joints and contact location. Building upon this unified description, a kinematic-aware planner model is proposed to generate precise 3D manipulation waypoints via a designed kinematic-aware chain-of-thoughts prompting method. Our evaluation spanned 48 instances across 16 distinct categories, revealing that our framework not only outperforms traditional methods on 8 seen categories but also shows a powerful zero-shot capability for 8 unseen articulated object categories. Moreover, the real-world experiments on 7 different object categories prove our framework's adaptability in practical scenarios. Code is released at https://github.com/GeWu-Lab/LLM_articulated_object_manipulation/tree/main.

Yake Wei, Ruoxuan Feng, Zihe Wang et al.

One primary topic of multimodal learning is to jointly incorporate heterogeneous information from different modalities. However most models often suffer from unsatisfactory multimodal cooperation which cannot jointly utilize all modalities well. Some methods are proposed to identify and enhance the worse learnt modality but they are often hard to provide the fine-grained observation of multimodal cooperation at sample-level with theoretical support. Hence it is essential to reasonably observe and improve the fine-grained cooperation between modalities especially when facing realistic scenarios where the modality discrepancy could vary across different samples. To this end we introduce a sample-level modality valuation metric to evaluate the contribution of each modality for each sample. Via modality valuation we observe that modality discrepancy indeed could be different at sample-level beyond the global contribution discrepancy at dataset-level. We further analyze this issue and improve cooperation between modalities at sample-level by enhancing the discriminative ability of low-contributing modalities in a targeted manner. Overall our methods reasonably observe the fine-grained uni-modal contribution and achieve considerable improvement. The source code and dataset are available at https://github.com/GeWu-Lab/Valuate-and-Enhance-Multimodal-Cooperation.

Guangyao Li, Wenxuan Hou, Di Hu

Audio-Visual Question Answering (AVQA) task aims to answer questions about different visual objects, sounds, and their associations in videos. Such naturally multi-modal videos are composed of rich and complex dynamic audio-visual components, where most of which could be unrelated to the given questions, or even play as interference in answering the content of interest. Oppositely, only focusing on the question-aware audio-visual content could get rid of influence, meanwhile enabling the model to answer more efficiently. In this paper, we propose a Progressive Spatio-Temporal Perception Network (PSTP-Net), which contains three modules that progressively identify key spatio-temporal regions w.r.t. questions. Specifically, a temporal segment selection module is first introduced to select the most relevant audio-visual segments related to the given question. Then, a spatial region selection module is utilized to choose the most relevant regions associated with the question from the selected temporal segments. To further refine the selection of features, an audio-guided visual attention module is employed to perceive the association between auido and selected spatial regions. Finally, the spatio-temporal features from these modules are integrated for answering the question. Extensive experimental results on the public MUSIC-AVQA and AVQA datasets provide compelling evidence of the effectiveness and efficiency of PSTP-Net. Code is available at: \href{https://github.com/GeWu-Lab/PSTP-Net}{https://github.com/GeWu-Lab/PSTP-Net}

Shihan Wu, Xuecheng Liu, Shaoxuan Xie et al.

Despite the critical role of bimanual manipulation in endowing robots with human-like dexterity, large-scale and diverse datasets remain scarce due to the significant hardware heterogeneity across bimanual robotic platforms. To bridge this gap, we introduce RoboCOIN, a large-scale multi-embodiment bimanual manipulation dataset comprising over 180,000 demonstrations collected from 15 distinct robotic platforms. Spanning 16 diverse environments-including residential, commercial, and industrial settings-the dataset features 421 bimanual tasks systematically categorized by 39 bimanual collaboration actions and 432 objects. A key innovation of our work is the hierarchical capability pyramid, which provides granular annotations ranging from trajectory-level concepts to segment-level subtasks and frame-level kinematics. Furthermore, we present CoRobot, an efficient data processing pipeline powered by the Robot Trajectory Markup Language (RTML), designed to facilitate quality assessment, automated annotation, and unified multi-embodiment and data management. Extensive experiments demonstrate the effectiveness of RoboCOIN in enhancing the performance of various bimanual manipulation models across a wide spectrum of robotic embodiments. The entire dataset and codebase are fully open-sourced, providing a valuable resource for advancing research in bimanual and multi-embodiment manipulation.

Yake Wei, Siwei Li, Ruoxuan Feng et al.

To overcome the imbalanced multimodal learning problem, where models prefer the training of specific modalities, existing methods propose to control the training of uni-modal encoders from different perspectives, taking the inter-modal performance discrepancy as the basis. However, the intrinsic limitation of modality capacity is ignored. The scarcely informative modalities can be recognized as ``worse-learnt'' ones, which could force the model to memorize more noise, counterproductively affecting the multimodal model ability. Moreover, the current modality modulation methods narrowly concentrate on selected worse-learnt modalities, even suppressing the training of others. Hence, it is essential to consider the intrinsic limitation of modality capacity and take all modalities into account during balancing. To this end, we propose the Diagnosing \& Re-learning method. The learning state of each modality is firstly estimated based on the separability of its uni-modal representation space, and then used to softly re-initialize the corresponding uni-modal encoder. In this way, the over-emphasizing of scarcely informative modalities is avoided. In addition, encoders of worse-learnt modalities are enhanced, simultaneously avoiding the over-training of other modalities. Accordingly, multimodal learning is effectively balanced and enhanced. Experiments covering multiple types of modalities and multimodal frameworks demonstrate the superior performance of our simple-yet-effective method for balanced multimodal learning. The source code and dataset are available at \url{https://github.com/GeWu-Lab/Diagnosing_Relearning_ECCV2024}.

Hongpeng Lin, Ludan Ruan, Wenke Xia et al.

To facilitate the research on intelligent and human-like chatbots with multi-modal context, we introduce a new video-based multi-modal dialogue dataset, called TikTalk. We collect 38K videos from a popular video-sharing platform, along with 367K conversations posted by users beneath them. Users engage in spontaneous conversations based on their multi-modal experiences from watching videos, which helps recreate real-world chitchat context. Compared to previous multi-modal dialogue datasets, the richer context types in TikTalk lead to more diverse conversations, but also increase the difficulty in capturing human interests from intricate multi-modal information to generate personalized responses. Moreover, external knowledge is more frequently evoked in our dataset. These facts reveal new challenges for multi-modal dialogue models. We quantitatively demonstrate the characteristics of TikTalk, propose a video-based multi-modal chitchat task, and evaluate several dialogue baselines. Experimental results indicate that the models incorporating large language models (LLM) can generate more diverse responses, while the model utilizing knowledge graphs to introduce external knowledge performs the best overall. Furthermore, no existing model can solve all the above challenges well. There is still a large room for future improvements, even for LLM with visual extensions. Our dataset is available at \url{https://ruc-aimind.github.io/projects/TikTalk/}.

Guangyao Li, Henghui Du, Di Hu

The Audio Visual Question Answering (AVQA) task aims to answer questions related to various visual objects, sounds, and their interactions in videos. Such naturally multimodal videos contain rich and complex dynamic audio-visual components, with only a portion of them closely related to the given questions. Hence, effectively perceiving audio-visual cues relevant to the given questions is crucial for correctly answering them. In this paper, we propose a Temporal-Spatial Perception Model (TSPM), which aims to empower the model to perceive key visual and auditory cues related to the questions. Specifically, considering the challenge of aligning non-declarative questions and visual representations into the same semantic space using visual-language pretrained models, we construct declarative sentence prompts derived from the question template, to assist the temporal perception module in better identifying critical segments relevant to the questions. Subsequently, a spatial perception module is designed to merge visual tokens from selected segments to highlight key latent targets, followed by cross-modal interaction with audio to perceive potential sound-aware areas. Finally, the significant temporal-spatial cues from these modules are integrated to answer the question. Extensive experiments on multiple AVQA benchmarks demonstrate that our framework excels not only in understanding audio-visual scenes but also in answering complex questions effectively. Code is available at https://github.com/GeWu-Lab/TSPM.

Ziyun Li, Jona Otholt, Ben Dai et al.

Novel class discovery (NCD) aims to infer novel categories in an unlabeled dataset leveraging prior knowledge of a labeled set comprising disjoint but related classes. Existing research focuses primarily on utilizing the labeled set at the methodological level, with less emphasis on the analysis of the labeled set itself. Thus, in this paper, we rethink novel class discovery from the labeled set and focus on two core questions: (i) Given a specific unlabeled set, what kind of labeled set can best support novel class discovery? (ii) A fundamental premise of NCD is that the labeled set must be related to the unlabeled set, but how can we measure this relation? For (i), we propose and substantiate the hypothesis that NCD could benefit more from a labeled set with a large degree of semantic similarity to the unlabeled set. Specifically, we establish an extensive and large-scale benchmark with varying degrees of semantic similarity between labeled/unlabeled datasets on ImageNet by leveraging its hierarchical class structure. As a sharp contrast, the existing NCD benchmarks are developed based on labeled sets with different number of categories and images, and completely ignore the semantic relation. For (ii), we introduce a mathematical definition for quantifying the semantic similarity between labeled and unlabeled sets. In addition, we use this metric to confirm the validity of our proposed benchmark and demonstrate that it highly correlates with NCD performance. Furthermore, without quantitative analysis, previous works commonly believe that label information is always beneficial. However, counterintuitively, our experimental results show that using labels may lead to sub-optimal outcomes in low-similarity settings.

Xinchi Zhou, Dongzhan Zhou, Wanli Ouyang et al.

Recent years have witnessed the success of deep learning on the visual sound separation task. However, existing works follow similar settings where the training and testing datasets share the same musical instrument categories, which to some extent limits the versatility of this task. In this work, we focus on a more general and challenging scenario, namely the separation of unknown musical instruments, where the categories in training and testing phases have no overlap with each other. To tackle this new setting, we propose the Separation-with-Consistency (SeCo) framework, which can accomplish the separation on unknown categories by exploiting the consistency constraints. Furthermore, to capture richer characteristics of the novel melodies, we devise an online matching strategy, which can bring stable enhancements with no cost of extra parameters. Experiments demonstrate that our SeCo framework exhibits strong adaptation ability on the novel musical categories and outperforms the baseline methods by a significant margin.

Yake Wei, Di Hu, Yapeng Tian et al.

Sight and hearing are two senses that play a vital role in human communication and scene understanding. To mimic human perception ability, audio-visual learning, aimed at developing computational approaches to learn from both audio and visual modalities, has been a flourishing field in recent years. A comprehensive survey that can systematically organize and analyze studies of the audio-visual field is expected. Starting from the analysis of audio-visual cognition foundations, we introduce several key findings that have inspired our computational studies. Then, we systematically review the recent audio-visual learning studies and divide them into three categories: audio-visual boosting, cross-modal perception and audio-visual collaboration. Through our analysis, we discover that, the consistency of audio-visual data across semantic, spatial and temporal support the above studies. To revisit the current development of the audio-visual learning field from a more macro view, we further propose a new perspective on audio-visual scene understanding, then discuss and analyze the feasible future direction of the audio-visual learning area. Overall, this survey reviews and outlooks the current audio-visual learning field from different aspects. We hope it can provide researchers with a better understanding of this area. A website including constantly-updated survey is released: \url{https://gewu-lab.github.io/audio-visual-learning/}.

Andong Deng, Xingjian Li, Di Hu et al.

Pre-training has been a popular learning paradigm in deep learning era, especially in annotation-insufficient scenario. Better ImageNet pre-trained models have been demonstrated, from the perspective of architecture, by previous research to have better transferability to downstream tasks. However, in this paper, we found that during the same pre-training process, models at middle epochs, which is inadequately pre-trained, can outperform fully trained models when used as feature extractors (FE), while the fine-tuning (FT) performance still grows with the source performance. This reveals that there is not a solid positive correlation between top-1 accuracy on ImageNet and the transferring result on target data. Based on the contradictory phenomenon between FE and FT that better feature extractor fails to be fine-tuned better accordingly, we conduct comprehensive analyses on features before softmax layer to provide insightful explanations. Our discoveries suggest that, during pre-training, models tend to first learn spectral components corresponding to large singular values and the residual components contribute more when fine-tuning.

Yaoting Wang, Peiwen Sun, Yuanchao Li et al.

The Audio-Visual Segmentation (AVS) task aims to segment sounding objects in the visual space using audio cues. However, in this work, it is recognized that previous AVS methods show a heavy reliance on detrimental segmentation preferences related to audible objects, rather than precise audio guidance. We argue that the primary reason is that audio lacks robust semantics compared to vision, especially in multi-source sounding scenes, resulting in weak audio guidance over the visual space. Motivated by the the fact that text modality is well explored and contains rich abstract semantics, we propose leveraging text cues from the visual scene to enhance audio guidance with the semantics inherent in text. Our approach begins by obtaining scene descriptions through an off-the-shelf image captioner and prompting a frozen large language model to deduce potential sounding objects as text cues. Subsequently, we introduce a novel semantics-driven audio modeling module with a dynamic mask to integrate audio features with text cues, leading to representative sounding object features. These features not only encompass audio cues but also possess vivid semantics, providing clearer guidance in the visual space. Experimental results on AVS benchmarks validate that our method exhibits enhanced sensitivity to audio when aided by text cues, achieving highly competitive performance on all three subsets. Project page: \href{https://github.com/GeWu-Lab/Sounding-Object-Segmentation-Preference}{https://github.com/GeWu-Lab/Sounding-Object-Segmentation-Preference}

Yu Miao, Zequn Yang, Yake Wei et al. · pku

In different multimodal scenarios, it needs to integrate and utilize information across modalities in a specific way based on the demands of the task. Different integration ways between modalities are referred to as "multimodal interaction". How well a model handles various multimodal interactions largely characterizes its multimodal ability. In this paper, we introduce MIBench, a comprehensive benchmark designed to evaluate the multimodal interaction capabilities of Large Multimodal Models (LMMs), which formulates each instance as a (con_v , con_t, task) triplet with contexts from vision and text, necessitating that LMMs employ correct forms of multimodal interaction to effectively complete the task. MIBench assesses models from three key aspects: the ability to source information from vision-centric or text-centric cues, and the ability to generate new information from their joint synergy. Each interaction capability is evaluated hierarchically across three cognitive levels: Recognition, Understanding, and Reasoning. MIBench comprises over 10,000 vision-text context pairs spanning 32 distinct tasks. Evaluation of state-of-the-art LMMs show that: (1) LMMs' ability on multimodal interaction remains constrained, despite the scaling of model parameters and training data; (2) they are easily distracted by textual modalities when processing vision information; (3) they mostly possess a basic capacity for multimodal synergy; and (4) natively trained multimodal models show noticeable deficits in fundamental interaction ability. We expect that these observations can serve as a reference for developing LMMs with more enhanced multimodal ability in the future.

Di Hu, Kun Li, Haojie Rao et al.

Accurate prediction of molecular properties underpins drug discovery and material design, yet even state-of-the-art models remain vulnerable to localized failure modes that aggregate metrics cannot detect. The places where molecular similarity should be most helpful are also places where standard evaluation can be most misleading. Property cliffs expose this gap: structurally similar molecules can still differ sharply in target property, so models with competitive overall performance may fail in high-risk local neighborhoods. To expose and mitigate this failure mode, CliffSplit, a cliff-aware evaluation protocol that constructs locally supported, cliff-exposed test cases, and CliffLoss, a model-agnostic train-only mitigation mechanism for cliff-sensitive errors, are introduced. Experiments on three QM9 targets and three MoleculeNet tasks across five backbones show that CliffSplit reveals at least 15% higher error in cliff-heavy QM9 regions, while CliffLoss reduces the cliff-to-smooth error gap by up to 30% on Lipophilicity and improves overall MAE by 9.7%. Together, these results turn molecular similarity failure from a descriptive anomaly into a benchmarked evaluation problem for molecular machine learning. The code is available at https://anonymous.4open.science/r/Cliff_Loss.

Jingxian Lu, Wenke Xia, Dong Wang et al.

Online Imitation Learning struggles with the gap between extensive online exploration space and limited expert trajectories, hindering efficient exploration due to inaccurate reward estimation. Inspired by the findings from cognitive neuroscience, we hypothesize that an agent could estimate precise task-aware reward for efficient online exploration, through decomposing the target task into the objectives of "what to do" and the mechanisms of "how to do". In this work, we introduce the hybrid Key-state guided Online Imitation (KOI) learning method, which leverages the integration of semantic and motion key states as guidance for reward estimation. Initially, we utilize visual-language models to extract semantic key states from expert trajectory, indicating the objectives of "what to do". Within the intervals between semantic key states, optical flow is employed to capture motion key states to understand the mechanisms of "how to do". By integrating a thorough grasp of hybrid key states, we refine the trajectory-matching reward computation, accelerating online imitation learning with task-aware exploration. We evaluate not only the success rate of the tasks in the Meta-World and LIBERO environments, but also the trend of variance during online imitation learning, proving that our method is more sample efficient. We also conduct real-world robotic manipulation experiments to validate the efficacy of our method, demonstrating the practical applicability of our KOI method. Videos and code are available at https://gewu-lab.github.io/Keystate_Online_Imitation/.

Yaoting Wang, Weisong Liu, Guangyao Li et al.

Never having seen an object and heard its sound simultaneously, can the model still accurately localize its visual position from the input audio? In this work, we concentrate on the Audio-Visual Localization and Segmentation tasks but under the demanding zero-shot and few-shot scenarios. To achieve this goal, different from existing approaches that mostly employ the encoder-fusion-decoder paradigm to decode localization information from the fused audio-visual feature, we introduce the encoder-prompt-decoder paradigm, aiming to better fit the data scarcity and varying data distribution dilemmas with the help of abundant knowledge from pre-trained models. Specifically, we first propose to construct Semantic-aware Audio Prompt (SAP) to help the visual foundation model focus on sounding objects, meanwhile, the semantic gap between the visual and audio modalities is also encouraged to shrink. Then, we develop a Correlation Adapter (ColA) to keep minimal training efforts as well as maintain adequate knowledge of the visual foundation model. By equipping with these means, extensive experiments demonstrate that this new paradigm outperforms other fusion-based methods in both the unseen class and cross-dataset settings. We hope that our work can further promote the generalization study of Audio-Visual Localization and Segmentation in practical application scenarios.

Yaoting Wang, Peiwen Sun, Dongzhan Zhou et al.

Traditional reference segmentation tasks have predominantly focused on silent visual scenes, neglecting the integral role of multimodal perception and interaction in human experiences. In this work, we introduce a novel task called Reference Audio-Visual Segmentation (Ref-AVS), which seeks to segment objects within the visual domain based on expressions containing multimodal cues. Such expressions are articulated in natural language forms but are enriched with multimodal cues, including audio and visual descriptions. To facilitate this research, we construct the first Ref-AVS benchmark, which provides pixel-level annotations for objects described in corresponding multimodal-cue expressions. To tackle the Ref-AVS task, we propose a new method that adequately utilizes multimodal cues to offer precise segmentation guidance. Finally, we conduct quantitative and qualitative experiments on three test subsets to compare our approach with existing methods from related tasks. The results demonstrate the effectiveness of our method, highlighting its capability to precisely segment objects using multimodal-cue expressions. Dataset is available at \href{https://gewu-lab.github.io/Ref-AVS}{https://gewu-lab.github.io/Ref-AVS}.

Ruoxuan Feng, Di Hu, Wenke Ma et al.

Humans possess a remarkable talent for flexibly alternating to different senses when interacting with the environment. Picture a chef skillfully gauging the timing of ingredient additions and controlling the heat according to the colors, sounds, and aromas, seamlessly navigating through every stage of the complex cooking process. This ability is founded upon a thorough comprehension of task stages, as achieving the sub-goal within each stage can necessitate the utilization of different senses. In order to endow robots with similar ability, we incorporate the task stages divided by sub-goals into the imitation learning process to accordingly guide dynamic multi-sensory fusion. We propose MS-Bot, a stage-guided dynamic multi-sensory fusion method with coarse-to-fine stage understanding, which dynamically adjusts the priority of modalities based on the fine-grained state within the predicted current stage. We train a robot system equipped with visual, auditory, and tactile sensors to accomplish challenging robotic manipulation tasks: pouring and peg insertion with keyway. Experimental results indicate that our approach enables more effective and explainable dynamic fusion, aligning more closely with the human fusion process than existing methods.

Juncheng Ma, Peiwen Sun, Yaoting Wang et al.

Audio-Visual Segmentation (AVS) aims to achieve pixel-level localization of sound sources in videos, while Audio-Visual Semantic Segmentation (AVSS), as an extension of AVS, further pursues semantic understanding of audio-visual scenes. However, since the AVSS task requires the establishment of audio-visual correspondence and semantic understanding simultaneously, we observe that previous methods have struggled to handle this mashup of objectives in end-to-end training, resulting in insufficient learning and sub-optimization. Therefore, we propose a two-stage training strategy called \textit{Stepping Stones}, which decomposes the AVSS task into two simple subtasks from localization to semantic understanding, which are fully optimized in each stage to achieve step-by-step global optimization. This training strategy has also proved its generalization and effectiveness on existing methods. To further improve the performance of AVS tasks, we propose a novel framework Adaptive Audio Visual Segmentation, in which we incorporate an adaptive audio query generator and integrate masked attention into the transformer decoder, facilitating the adaptive fusion of visual and audio features. Extensive experiments demonstrate that our methods achieve state-of-the-art results on all three AVS benchmarks. The project homepage can be accessed at https://gewu-lab.github.io/stepping_stones/.

Wenke Xia, Xu Zhao, Xincheng Pang et al.

The imbalance problem is widespread in the field of machine learning, which also exists in multimodal learning areas caused by the intrinsic discrepancy between modalities of samples. Recent works have attempted to solve the modality imbalance problem from algorithm perspective, however, they do not fully analyze the influence of modality bias in datasets. Concretely, existing multimodal datasets are usually collected under specific tasks, where one modality tends to perform better than other ones in most conditions. In this work, to comprehensively explore the influence of modality bias, we first split existing datasets into different subsets by estimating sample-wise modality discrepancy. We surprisingly find that: the multimodal models with existing imbalance algorithms consistently perform worse than the unimodal one on specific subsets, in accordance with the modality bias. To further explore the influence of modality bias and analyze the effectiveness of existing imbalance algorithms, we build a balanced audiovisual dataset, with uniformly distributed modality discrepancy over the whole dataset. We then conduct extensive experiments to re-evaluate existing imbalance algorithms and draw some interesting findings: existing algorithms only provide a compromise between modalities and suffer from the large modality discrepancy of samples. We hope that these findings could facilitate future research on the modality imbalance problem.

Yake Wei, Di Hu, Henghui Du et al.

Multimodal learning is expected to boost model performance by integrating information from different modalities. However, its potential is not fully exploited because the widely-used joint training strategy, which has a uniform objective for all modalities, leads to imbalanced and under-optimized uni-modal representations. Specifically, we point out that there often exists modality with more discriminative information, e.g., vision of playing football and sound of blowing wind. They could dominate the joint training process, resulting in other modalities being significantly under-optimized. To alleviate this problem, we first analyze the under-optimized phenomenon from both the feed-forward and the back-propagation stages during optimization. Then, On-the-fly Prediction Modulation (OPM) and On-the-fly Gradient Modulation (OGM) strategies are proposed to modulate the optimization of each modality, by monitoring the discriminative discrepancy between modalities during training. Concretely, OPM weakens the influence of the dominant modality by dropping its feature with dynamical probability in the feed-forward stage, while OGM mitigates its gradient in the back-propagation stage. In experiments, our methods demonstrate considerable improvement across a variety of multimodal tasks. These simple yet effective strategies not only enhance performance in vanilla and task-oriented multimodal models, but also in more complex multimodal tasks, showcasing their effectiveness and flexibility. The source code is available at \url{https://github.com/GeWu-Lab/BML_TPAMI2024}.

Yingzi Fan, Longfei Han, Yue Zhang et al.

Both visual and auditory information are valuable to determine the salient regions in videos. Deep convolution neural networks (CNN) showcase strong capacity in coping with the audio-visual saliency prediction task. Due to various factors such as shooting scenes and weather, there often exists moderate distribution discrepancy between source training data and target testing data. The domain discrepancy induces to performance degradation on target testing data for CNN models. This paper makes an early attempt to tackle the unsupervised domain adaptation problem for audio-visual saliency prediction. We propose a dual domain-adversarial learning algorithm to mitigate the domain discrepancy between source and target data. First, a specific domain discrimination branch is built up for aligning the auditory feature distributions. Then, those auditory features are fused into the visual features through a cross-modal self-attention module. The other domain discrimination branch is devised to reduce the domain discrepancy of visual features and audio-visual correlations implied by the fused audio-visual features. Experiments on public benchmarks demonstrate that our method can relieve the performance degradation caused by domain discrepancy.

Peiwen Sun, Honggang Zhang, Di Hu

Community researchers have developed a range of advanced audio-visual segmentation models aimed at improving the quality of sounding objects' masks. While masks created by these models may initially appear plausible, they occasionally exhibit anomalies with incorrect grounding logic. We attribute this to real-world inherent preferences and distributions as a simpler signal for learning than the complex audio-visual grounding, which leads to the disregard of important modality information. Generally, the anomalous phenomena are often complex and cannot be directly observed systematically. In this study, we made a pioneering effort with the proper synthetic data to categorize and analyze phenomena as two types "audio priming bias" and "visual prior" according to the source of anomalies. For audio priming bias, to enhance audio sensitivity to different intensities and semantics, a perception module specifically for audio perceives the latent semantic information and incorporates information into a limited set of queries, namely active queries. Moreover, the interaction mechanism related to such active queries in the transformer decoder is customized to adapt to the need for interaction regulating among audio semantics. For visual prior, multiple contrastive training strategies are explored to optimize the model by incorporating a biased branch, without even changing the structure of the model. During experiments, observation demonstrates the presence and the impact that has been produced by the biases of the existing model. Finally, through experimental evaluation of AVS benchmarks, we demonstrate the effectiveness of our methods in handling both types of biases, achieving competitive performance across all three subsets.

Ruoxuan Feng, Wenke Xia, Di Hu

Pre-training technique has gained tremendous success in enhancing model performance on various tasks, but found to perform worse than training from scratch in some uni-modal situations. This inspires us to think: are the pre-trained models always effective in the more complex multi-modal scenario, especially for the heterogeneous modalities such as audio and visual ones? We find that the answer is No. Specifically, we explore the effects of pre-trained models on two audio-visual learning scenarios: cross-modal initialization and multi-modal joint learning. When cross-modal initialization is applied, the phenomena of "dead channel" caused by abnormal Batchnorm parameters hinders the utilization of model capacity. Thus, we propose Adaptive Batchnorm Re-initialization (ABRi) to better exploit the capacity of pre-trained models for target tasks. In multi-modal joint learning, we find a strong pre-trained uni-modal encoder would bring negative effects on the encoder of another modality. To alleviate such problem, we introduce a two-stage Fusion Tuning strategy, taking better advantage of the pre-trained knowledge while making the uni-modal encoders cooperate with an adaptive masking method. The experiment results show that our methods could further exploit pre-trained models' potential and boost performance in audio-visual learning.

Henghui Du, Guangyao Li, Chang Zhou et al.

In recent years, numerous tasks have been proposed to encourage model to develop specified capability in understanding audio-visual scene, primarily categorized into temporal localization, spatial localization, spatio-temporal reasoning, and pixel-level understanding. Instead, human possesses a unified understanding ability for diversified tasks. Therefore, designing an audio-visual model with general capability to unify these tasks is of great value. However, simply joint training for all tasks can lead to interference due to the heterogeneity of audiovisual data and complex relationship among tasks. We argue that this problem can be solved through explicit cooperation among tasks. To achieve this goal, we propose a unified learning method which achieves explicit inter-task cooperation from both the perspectives of data and model thoroughly. Specifically, considering the labels of existing datasets are simple words, we carefully refine these datasets and construct an Audio-Visual Unified Instruction-tuning dataset with Explicit reasoning process (AV-UIE), which clarifies the cooperative relationship among tasks. Subsequently, to facilitate concrete cooperation in learning stage, an interaction-aware LoRA structure with multiple LoRA heads is designed to learn different aspects of audiovisual data interaction. By unifying the explicit cooperation across the data and model aspect, our method not only surpasses existing unified audio-visual model on multiple tasks, but also outperforms most specialized models for certain tasks. Furthermore, we also visualize the process of explicit cooperation and surprisingly find that each LoRA head has certain audio-visual understanding ability. Code and dataset: https://github.com/GeWu-Lab/Crab

Wenke Xia, Ruoxuan Feng, Dong Wang et al.

Building a generalizable self-correction system is crucial for robots to recover from failures. Despite advancements in Multimodal Large Language Models (MLLMs) that empower robots with semantic reflection ability for failure, translating semantic reflection into how to correct fine-grained robotic actions remains a significant challenge. To address this gap, we build the Phoenix framework, which leverages motion instruction as a bridge to connect high-level semantic reflection with low-level robotic action correction. In this motion-based self-reflection framework, we start with a dual-process motion adjustment mechanism with MLLMs to translate the semantic reflection into coarse-grained motion instruction adjustment. To leverage this motion instruction for guiding how to correct fine-grained robotic actions, a multi-task motion-conditioned diffusion policy is proposed to integrate visual observations for high-frequency robotic action correction. By combining these two models, we could shift the demand for generalization capability from the low-level manipulation policy to the MLLMs-driven motion adjustment model and facilitate precise, fine-grained robotic action correction. Utilizing this framework, we further develop a lifelong learning method to automatically improve the model's capability from interactions with dynamic environments. The experiments conducted in both the RoboMimic simulation and real-world scenarios prove the superior generalization and robustness of our framework across a variety of manipulation tasks. Our code is released at \href{https://github.com/GeWu-Lab/Motion-based-Self-Reflection-Framework}{https://github.com/GeWu-Lab/Motion-based-Self-Reflection-Framework}.

Shaoxuan Xu, Menglu Cui, Chengxiang Huang et al.

Multimodal learning has gained attention for its capacity to integrate information from different modalities. However, it is often hindered by the multimodal imbalance problem, where certain modality dominates while others remain underutilized. Although recent studies have proposed various methods to alleviate this problem, they lack comprehensive and fair comparisons. In this paper, we systematically categorize various mainstream multimodal imbalance algorithms into four groups based on the strategies they employ to mitigate imbalance. To facilitate a comprehensive evaluation of these methods, we introduce BalanceBenchmark, a benchmark including multiple widely used multidimensional datasets and evaluation metrics from three perspectives: performance, imbalance degree, and complexity. To ensure fair comparisons, we have developed a modular and extensible toolkit that standardizes the experimental workflow across different methods. Based on the experiments using BalanceBenchmark, we have identified several key insights into the characteristics and advantages of different method groups in terms of performance, balance degree and computational complexity. We expect such analysis could inspire more efficient approaches to address the imbalance problem in the future, as well as foundation models. The code of the toolkit is available at https://github.com/GeWu-Lab/BalanceBenchmark.

Xi Lu, Di Hu, An T. Nguyen et al.

Patient-controlled data-sharing systems are increasingly promoted as a way to empower patients with greater autonomy over their health data. Yet it remains unclear how different stakeholders, especially patients and health system leaders, perceive the benefits and challenges of enabling granular control over the sharing of de-identified medical data for research. To address this gap, we developed a high-fidelity prototype of a patient-controlled, web-based consent platform and conducted a two-phase mixed-methods study:semi-structured interviews with 16 health system leaders and a survey with 523 patient participants. While both groups appreciated the potential of such a platform to enhance transparency and autonomy, their views diverged in meaningful ways. Leaders viewed transparency and granular control through the lens of informed consent and institutional ethics, whereas patients interpreted these factors as safeguards against potential risks and uncertainties. Our findings underscore critical tensions such as individual control and research integrity. We offer design implications for building trustworthy, context-aware systems that support flexible granularity, provide ongoing benefit-centered transparency, and adapt to diverse literacy and privacy needs.

Chengxiang Huang, Yake Wei, Zequn Yang et al.

Sensory training during the early ages is vital for human development. Inspired by this cognitive phenomenon, we observe that the early training stage is also important for the multimodal learning process, where dataset information is rapidly acquired. We refer to this stage as the prime learning window. However, based on our observation, this prime learning window in multimodal learning is often dominated by information-sufficient modalities, which in turn suppresses the information acquisition of information-insufficient modalities. To address this issue, we propose Information Acquisition Regulation (InfoReg), a method designed to balance information acquisition among modalities. Specifically, InfoReg slows down the information acquisition process of information-sufficient modalities during the prime learning window, which could promote information acquisition of information-insufficient modalities. This regulation enables a more balanced learning process and improves the overall performance of the multimodal network. Experiments show that InfoReg outperforms related multimodal imbalanced methods across various datasets, achieving superior model performance. The code is available at https://github.com/GeWu-Lab/InfoReg_CVPR2025.

Ruoxuan Feng, Yuxuan Zhou, Siyu Mei et al.

Real-world contact-rich manipulation demands robots to perceive temporal tactile feedback, capture subtle surface deformations, and reason about object properties as well as force dynamics. Although optical tactile sensors are uniquely capable of providing such rich information, existing tactile datasets and models remain limited. These resources primarily focus on object-level attributes (e.g., material) while largely overlooking fine-grained tactile temporal dynamics during physical interactions. We consider that advancing dynamic tactile perception requires a systematic hierarchy of dynamic perception capabilities to guide both data collection and model design. To address the lack of tactile data with rich dynamic information, we present ToucHD, a large-scale hierarchical tactile dataset spanning tactile atomic actions, real-world manipulations, and touch-force paired data. Beyond scale, ToucHD establishes a comprehensive tactile dynamic data ecosystem that explicitly supports hierarchical perception capabilities from the data perspective. Building on it, we propose AnyTouch 2, a general tactile representation learning framework for diverse optical tactile sensors that unifies object-level understanding with fine-grained, force-aware dynamic perception. The framework captures both pixel-level and action-specific deformations across frames, while explicitly modeling physical force dynamics, thereby learning multi-level dynamic perception capabilities from the model perspective. We evaluate our model on benchmarks that covers static object properties and dynamic physical attributes, as well as real-world manipulation tasks spanning multiple tiers of dynamic perception capabilities-from basic object-level understanding to force-aware dexterous manipulation. Experimental results demonstrate consistent and strong performance across sensors and tasks.

Wenke Xia, Yichu Yang, Hongtao Wu et al.

Establishing a reliable and iteratively refined robotic system is essential for deploying real-world applications. While Vision-Language-Action (VLA) models are widely recognized as the foundation model for such robotic deployment, their reliance on offline expert demonstrations critically limits their capacity for post-deployment refinement. To mitigate this limitation, we introduce Action Preference Optimization (APO), a method designed to refine VLA models by human-assisted preference alignment gathered through interaction with environments. This method begins with a human-robot collaboration framework for reliable failure correction and interaction trajectory collection through human intervention. However, directly leveraging these interaction trajectories for preference optimization is non-trivial due to the challenges of irreversible robotic actions and token distribution mismatch. To solve this, APO proposes an adaptive reweighting algorithm with binary desirability signals derived from interaction, empowering VLA models effectively suppress failure-prone actions while enhancing corrective action adaptation. Ultimately, APO equips VLA models with the crucial capability to learn from failure, paving the way for their iterative refinement and reliable deployment in dynamic environments. The experiments conducted in simulation and real-world scenarios prove superior generalization and robustness of our human-assisted framework across a variety of manipulation tasks. We believe this work could bring insights for efficient and stable optimization of VLA models through human-robot collaboration. The code and dataset are released at https://github.com/GeWu-Lab/Action-Preference-Optimization

Zequn Yang, Hongfa Wang, Di Hu

Interactions between modalities -- redundancy, uniqueness, and synergy -- collectively determine the composition of multimodal information. Understanding these interactions is crucial for analyzing information dynamics in multimodal systems, yet their accurate sample-level quantification presents significant theoretical and computational challenges. To address this, we introduce the Lightweight Sample-wise Multimodal Interaction (LSMI) estimator, rigorously grounded in pointwise information theory. We first develop a redundancy estimation framework, employing an appropriate pointwise information measure to quantify this most decomposable and measurable interaction. Building upon this, we propose a general interaction estimation method that employs efficient entropy estimation, specifically tailored for sample-wise estimation in continuous distributions. Extensive experiments on synthetic and real-world datasets validate LSMI's precision and efficiency. Crucially, our sample-wise approach reveals fine-grained sample- and category-level dynamics within multimodal data, enabling practical applications such as redundancy-informed sample partitioning, targeted knowledge distillation, and interaction-aware model ensembling. The code is available at https://github.com/GeWu-Lab/LSMI_Estimator.

Ruotian Peng, Haiying He, Yake Wei et al.

High-quality image captions play a crucial role in improving the performance of cross-modal applications such as text-to-image generation, text-to-video generation, and text-image retrieval. To generate long-form, high-quality captions, many recent studies have employed multimodal large language models (MLLMs). However, current MLLMs often produce captions that lack fine-grained details or suffer from hallucinations, a challenge that persists in both open-source and closed-source models. Inspired by Feature-Integration theory, which suggests that attention must focus on specific regions to integrate visual information effectively, we propose a \textbf{divide-then-aggregate} strategy. Our method first divides the image into semantic and spatial patches to extract fine-grained details, enhancing the model's local perception of the image. These local details are then hierarchically aggregated to generate a comprehensive global description. To address hallucinations and inconsistencies in the generated captions, we apply a semantic-level filtering process during hierarchical aggregation. This training-free pipeline can be applied to both open-source models (LLaVA-1.5, LLaVA-1.6, Mini-Gemini) and closed-source models (Claude-3.5-Sonnet, GPT-4o, GLM-4V-Plus). Extensive experiments demonstrate that our method generates more detailed, reliable captions, advancing multimodal description generation without requiring model retraining. The source code are available at https://github.com/GeWu-Lab/Patch-Matters

Haotian Ni, Yake Wei, Hang Liu et al.

Multimodal learning faces challenges in effectively fusing information from diverse modalities, especially when modality quality varies across samples. Dynamic fusion strategies, such as attention mechanism in Transformers, aim to address such challenge by adaptively emphasizing modalities based on the characteristics of input data. However, through amounts of carefully designed experiments, we surprisingly observed that the dynamic adaptability of widely-used self-attention models diminishes. Model tends to prefer one modality regardless of data characteristics. This bias triggers a self-reinforcing cycle that progressively overemphasizes the favored modality, widening the distribution gap in attention keys across modalities and deactivating attention mechanism's dynamic properties. To revive adaptability, we propose a simple yet effective method Rolling Query (RollingQ), which balances attention allocation by rotating the query to break the self-reinforcing cycle and mitigate the key distribution gap. Extensive experiments on various multimodal scenarios validate the effectiveness of RollingQ and the restoration of cooperation dynamics is pivotal for enhancing the broader capabilities of widely deployed multimodal Transformers. The source code is available at https://github.com/GeWu-Lab/RollingQ_ICML2025.

Dongnuan Cai, Henghui Du, Chang Zhou et al.

Developing Audio-Visual Large Language Models (AV-LLMs) for unified scene understanding is pivotal in multimodal intelligence. While instruction tuning enables pre-trained models with multi-task abilities, we observe that conventional multi-task unification methods often suffer from severe negative transfer, where nearly 55% of tasks degrade compared to single-task training. We attribute this phenomenon to audio-visual task heterogeneity, characterized by disparate task granularity and divergent capability demands, which lead to negative interference under joint training. To tackle this, we present Crab$^{+}$, a scalable and unified audio-visual scene understanding model that addresses task heterogeneity through explicit cooperation from both data and model perspectives. On the data side, we introduce AV-UIE v2, a comprehensive Audio-Visual Unified Instruction-tuning dataset with Explicit reasoning processes. It contains approximately 222K samples spanning 17 datasets and 7 tasks, enabling the model to capture cross-task relationships at different levels of granularity. On the model side, we design a unified interface to align heterogeneous task formulations, and propose Interaction-aware LoRA (I-LoRA), which explicitly models inter-task relationships via dynamic routing to coordinate distinct audio-visual interaction patterns, mitigating parameter interference. Extensive experiments show Crab$^{+}$ covers broader tasks than existing unified models while outperforming specialized models on various benchmarks. We successfully reverse the negative transfer trend, achieving positive transfer where multi-task learning surpasses single-task baselines in nearly 88% of tasks. These results hold across diverse AV-LLM paradigms and are validated through in-depth visualization, positioning Crab$^{+}$ as a robust step towards holistic audio-visual scene understanding.

Yiliang Zhou, Yawen Guo, Di Hu et al.

Ambient AI documentation systems generate clinical note drafts that clinicians frequently revise before signing off into electronic health records, yet how these edits alter hedging language remains unclear. We conducted paired analysis of clinician-edited portions of ambient AI drafts and final notes to examine (1) whether these edits change the prevalence of hedging language, (2) whether these edits exhibit a systematic shift toward greater certainty or uncertainty, and (3) whether these changes in hedging prevalence and directionality differ by ambient AI vendors and clinical specialties. Among 62,811 paired note sections, hedging terms were more often introduced into previously non-hedged text than removed from previously hedged text, and post-edit text contained more hedging mentions than pre-edit text. Directionality analyses showed a significant overall tendency toward greater uncertainty in hedging-related replacement edits. Vendor and specialty analyses revealed substantial heterogeneity in hedging prevalence, pre-to-post changes in hedging mentions, and directionality.

Henghui Du, Chunjie Zhang, Xi Chen et al.

Long Video Question-Answering (LVQA) presents a significant challenge for Multi-modal Large Language Models (MLLMs) due to immense context and overloaded information, which could also lead to prohibitive memory consumption. While existing methods attempt to address these issues by reducing visual tokens or extending model's context length, they may miss useful information or take considerable computation. In fact, when answering given questions, only a small amount of crucial information is required. Therefore, we propose an efficient question-aware memory mechanism, enabling MLLMs to recurrently seek these critical clues. Our approach, named VideoDetective, simplifies this task by iteratively processing video sub-segments. For each sub-segment, a question-aware compression strategy is employed by introducing a few special memory tokens to achieve purposefully compression. This allows models to effectively seek critical clues while reducing visual tokens. Then, due to history context could have a significant impact, we recurrently aggregate and store these memory tokens to update history context, which would be reused for subsequent sub-segments. Furthermore, to more effectively measure model's long video understanding ability, we introduce GLVC (Grounding Long Video Clues), a long video question-answering dataset, which features grounding critical and concrete clues scattered throughout entire videos. Experimental results demonstrate our method enables MLLMs with limited context length of 32K to efficiently process 100K tokens (3600 frames, an hour-long video sampled at 1fps), requiring only 2 minutes and 37GB GPU memory usage. Evaluation results across multiple long video benchmarks illustrate our method can more effectively seek critical clues from massive information.

Yutao Zhu, Kun Zhou, Kelong Mao et al.

Large language models (LLMs) have become the foundation of many applications, leveraging their extensive capabilities in processing and understanding natural language. While many open-source LLMs have been released with technical reports, the lack of training details hinders further research and development. This paper presents the development of YuLan, a series of open-source LLMs with $12$ billion parameters. The base model of YuLan is pre-trained on approximately $1.7$T tokens derived from a diverse corpus, including massive English, Chinese, and multilingual texts. We design a three-stage pre-training method to enhance YuLan's overall capabilities. Subsequent phases of training incorporate instruction-tuning and human alignment, employing a substantial volume of high-quality synthesized data. To facilitate the learning of complex and long-tail knowledge, we devise a curriculum-learning framework throughout across these stages, which helps LLMs learn knowledge in an easy-to-hard manner. YuLan's training is finished on Jan, 2024 and has achieved performance on par with state-of-the-art LLMs across various English and Chinese benchmarks. This paper outlines a comprehensive technical roadmap for developing LLMs from scratch. Our model and codes are available at https://github.com/RUC-GSAI/YuLan-Chat.

Jia Zeng, Qingwen Bu, Bangjun Wang et al.

Representation learning approaches for robotic manipulation have boomed in recent years. Due to the scarcity of in-domain robot data, prevailing methodologies tend to leverage large-scale human video datasets to extract generalizable features for visuomotor policy learning. Despite the progress achieved, prior endeavors disregard the interactive dynamics that capture behavior patterns and physical interaction during the manipulation process, resulting in an inadequate understanding of the relationship between objects and the environment. To this end, we propose a general pre-training pipeline that learns Manipulation by Predicting the Interaction (MPI) and enhances the visual representation.Given a pair of keyframes representing the initial and final states, along with language instructions, our algorithm predicts the transition frame and detects the interaction object, respectively. These two learning objectives achieve superior comprehension towards "how-to-interact" and "where-to-interact". We conduct a comprehensive evaluation of several challenging robotic tasks.The experimental results demonstrate that MPI exhibits remarkable improvement by 10% to 64% compared with previous state-of-the-art in real-world robot platforms as well as simulation environments. Code and checkpoints are publicly shared at https://github.com/OpenDriveLab/MPI.

Guangyao Li, Yixin Xu, Di Hu

Audio question answering (AQA), acting as a widely used proxy task to explore scene understanding, has got more attention. The AQA is challenging for it requires comprehensive temporal reasoning from different scales' events of an audio scene. However, existing methods mostly extend the structures of visual question answering task to audio ones in a simple pattern but may not perform well when perceiving a fine-grained audio scene. To this end, we present a Multi-scale Window Attention Fusion Model (MWAFM) consisting of an asynchronous hybrid attention module and a multi-scale window attention module. The former is designed to aggregate unimodal and cross-modal temporal contexts, while the latter captures sound events of varying lengths and their temporal dependencies for a more comprehensive understanding. Extensive experiments are conducted to demonstrate that the proposed MWAFM can effectively explore temporal information to facilitate AQA in the fine-grained scene.Code: https://github.com/GeWu-Lab/MWAFM

Xian Liu, Rui Qian, Hang Zhou et al.

The task of audio-visual sound source localization has been well studied under constrained scenes, where the audio recordings are clean. However, in real-world scenarios, audios are usually contaminated by off-screen sound and background noise. They will interfere with the procedure of identifying desired sources and building visual-sound connections, making previous studies non-applicable. In this work, we propose the Interference Eraser (IEr) framework, which tackles the problem of audio-visual sound source localization in the wild. The key idea is to eliminate the interference by redefining and carving discriminative audio representations. Specifically, we observe that the previous practice of learning only a single audio representation is insufficient due to the additive nature of audio signals. We thus extend the audio representation with our Audio-Instance-Identifier module, which clearly distinguishes sounding instances when audio signals of different volumes are unevenly mixed. Then we erase the influence of the audible but off-screen sounds and the silent but visible objects by a Cross-modal Referrer module with cross-modality distillation. Quantitative and qualitative evaluations demonstrate that our proposed framework achieves superior results on sound localization tasks, especially under real-world scenarios. Code is available at https://github.com/alvinliu0/Visual-Sound-Localization-in-the-Wild.

Konrad Heidler, Lichao Mou, Di Hu et al.

Many current deep learning approaches make extensive use of backbone networks pre-trained on large datasets like ImageNet, which are then fine-tuned to perform a certain task. In remote sensing, the lack of comparable large annotated datasets and the wide diversity of sensing platforms impedes similar developments. In order to contribute towards the availability of pre-trained backbone networks in remote sensing, we devise a self-supervised approach for pre-training deep neural networks. By exploiting the correspondence between geo-tagged audio recordings and remote sensing imagery, this is done in a completely label-free manner, eliminating the need for laborious manual annotation. For this purpose, we introduce the SoundingEarth dataset, which consists of co-located aerial imagery and audio samples all around the world. Using this dataset, we then pre-train ResNet models to map samples from both modalities into a common embedding space, which encourages the models to understand key properties of a scene that influence both visual and auditory appearance. To validate the usefulness of the proposed approach, we evaluate the transfer learning performance of pre-trained weights obtained against weights obtained through other means. By fine-tuning the models on a number of commonly used remote sensing datasets, we show that our approach outperforms existing pre-training strategies for remote sensing imagery. The dataset, code and pre-trained model weights will be available at https://github.com/khdlr/SoundingEarth.

Dong Wang, Di Hu, Xingjian Li et al.

Temporal relational modeling in video is essential for human action understanding, such as action recognition and action segmentation. Although Graph Convolution Networks (GCNs) have shown promising advantages in relation reasoning on many tasks, it is still a challenge to apply graph convolution networks on long video sequences effectively. The main reason is that large number of nodes (i.e., video frames) makes GCNs hard to capture and model temporal relations in videos. To tackle this problem, in this paper, we introduce an effective GCN module, Dilated Temporal Graph Reasoning Module (DTGRM), designed to model temporal relations and dependencies between video frames at various time spans. In particular, we capture and model temporal relations via constructing multi-level dilated temporal graphs where the nodes represent frames from different moments in video. Moreover, to enhance temporal reasoning ability of the proposed model, an auxiliary self-supervised task is proposed to encourage the dilated temporal graph reasoning module to find and correct wrong temporal relations in videos. Our DTGRM model outperforms state-of-the-art action segmentation models on three challenging datasets: 50Salads, Georgia Tech Egocentric Activities (GTEA), and the Breakfast dataset. The code is available at https://github.com/redwang/DTGRM.

Di Hu, Rui Qian, Minyue Jiang et al.

Discriminatively localizing sounding objects in cocktail-party, i.e., mixed sound scenes, is commonplace for humans, but still challenging for machines. In this paper, we propose a two-stage learning framework to perform self-supervised class-aware sounding object localization. First, we propose to learn robust object representations by aggregating the candidate sound localization results in the single source scenes. Then, class-aware object localization maps are generated in the cocktail-party scenarios by referring the pre-learned object knowledge, and the sounding objects are accordingly selected by matching audio and visual object category distributions, where the audiovisual consistency is viewed as the self-supervised signal. Experimental results in both realistic and synthesized cocktail-party videos demonstrate that our model is superior in filtering out silent objects and pointing out the location of sounding objects of different classes. Code is available at https://github.com/DTaoo/Discriminative-Sounding-Objects-Localization.

Rui Qian, Di Hu, Heinrich Dinkel et al.

How to visually localize multiple sound sources in unconstrained videos is a formidable problem, especially when lack of the pairwise sound-object annotations. To solve this problem, we develop a two-stage audiovisual learning framework that disentangles audio and visual representations of different categories from complex scenes, then performs cross-modal feature alignment in a coarse-to-fine manner. Our model achieves state-of-the-art results on public dataset of localization, as well as considerable performance on multi-source sound localization in complex scenes. We then employ the localization results for sound separation and obtain comparable performance to existing methods. These outcomes demonstrate our model's ability in effectively aligning sounds with specific visual sources. Code is available at https://github.com/shvdiwnkozbw/Multi-Source-Sound-Localization