Wei Huang

Tianyu Liu, Peng Zhang, Wei Huang et al.

Self-supervised sound source localization is usually challenged by the modality inconsistency. In recent studies, contrastive learning based strategies have shown promising to establish such a consistent correspondence between audio and sound sources in visual scenarios. Unfortunately, the insufficient attention to the heterogeneity influence in the different modality features still limits this scheme to be further improved, which also becomes the motivation of our work. In this study, an Induction Network is proposed to bridge the modality gap more effectively. By decoupling the gradients of visual and audio modalities, the discriminative visual representations of sound sources can be learned with the designed Induction Vector in a bootstrap manner, which also enables the audio modality to be aligned with the visual modality consistently. In addition to a visual weighted contrastive loss, an adaptive threshold selection strategy is introduced to enhance the robustness of the Induction Network. Substantial experiments conducted on SoundNet-Flickr and VGG-Sound Source datasets have demonstrated a superior performance compared to other state-of-the-art works in different challenging scenarios. The code is available at https://github.com/Tahy1/AVIN

Junwen Xiong, Peng Zhang, Chuanyue Li et al.

Video saliency prediction and detection are thriving research domains that enable computers to simulate the distribution of visual attention akin to how humans perceiving dynamic scenes. While many approaches have crafted task-specific training paradigms for either video saliency prediction or video salient object detection tasks, few attention has been devoted to devising a generalized saliency modeling framework that seamlessly bridges both these distinct tasks. In this study, we introduce the Unified Saliency Transformer (UniST) framework, which comprehensively utilizes the essential attributes of video saliency prediction and video salient object detection. In addition to extracting representations of frame sequences, a saliency-aware transformer is designed to learn the spatio-temporal representations at progressively increased resolutions, while incorporating effective cross-scale saliency information to produce a robust representation. Furthermore, a task-specific decoder is proposed to perform the final prediction for each task. To the best of our knowledge, this is the first work that explores designing a transformer structure for both saliency modeling tasks. Convincible experiments demonstrate that the proposed UniST achieves superior performance across seven challenging benchmarks for two tasks, and significantly outperforms the other state-of-the-art methods.

Li Wang, Wei Huang

Variational Auto-Encoders (VAEs) have emerged as powerful probabilistic models for generative tasks. However, their convergence properties have not been rigorously proven. The challenge of proving convergence is inherently difficult due to the highly non-convex nature of the training objective and the implementation of a Stochastic Neural Network (SNN) within VAE architectures. This paper addresses these challenges by characterizing the optimization trajectory of SNNs utilized in VAEs through the lens of Neural Tangent Kernel (NTK) techniques. These techniques govern the optimization and generalization behaviors of ultra-wide neural networks. We provide a mathematical proof of VAE convergence under mild assumptions, thus advancing the theoretical understanding of VAE optimization dynamics. Furthermore, we establish a novel connection between the optimization problem faced by over-parameterized SNNs and the Kernel Ridge Regression (KRR) problem. Our findings not only contribute to the theoretical foundation of VAEs but also open new avenues for investigating the optimization of generative models using advanced kernel methods. Our theoretical claims are verified by experimental simulations.

Huanyu Li, Zongyuan Li, Wei Huang et al.

Large language models (LLMs) such as ChatGPT o1, ChatGPT o3, and DeepSeek R1 have shown great potential in solving difficult problems. However, current LLM evaluation benchmarks are limited to one-step interactions. Some of the existing sequence decision-making environments, such as TextStarCraftII and LLM-PySC2, are too complicated and require hours of interaction to complete a game. In this paper, we introduce LLM-Cave, a benchmark and light environment for LLM reasoning and decision-making systems. This environment is a classic instance in the era of Symbolism. Artificial intelligence enables the agent to explore the environment and avoid potential losses by reasoning about nearby dangers using partial observable state information. In the experiment, we evaluated the sequential reasoning ability, decision-making performance and computational efficiency of mainstream large language models (LLMs) such as GPT-4o-mini, o1-mini, and DeepSeek-R1. Experiments show that while Deepseek-R1 achieved the highest success rate on complex reasoning tasks, smaller models like 4o-mini significantly narrowed the performance gap on challenges by employing Chain of Speculation and Planner-Critic strategies, at the expense of reduced computational efficiency. This indicates that structured, multi-step reasoning combined with an LLM-based feedback mechanism can substantially enhance an LLM's decision-making capabilities, providing a promising direction for improving reasoning in weaker models and suggesting a new reasoning-centered benchmark for LLM assessment. Our code is open-sourced in https://github.com/puleya1277/CaveEnv.

Yukang Chen, Wei Huang, Baifeng Shi et al.

We introduce a full-stack framework that scales up reasoning in vision-language models (VLMs) to long videos, leveraging reinforcement learning. We address the unique challenges of long video reasoning by integrating three critical components: (1) a large-scale dataset, LongVideo-Reason, comprising 104K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. In our experiments, LongVILA-R1-7B achieves strong performance on video benchmarks, reaching 65.1% and 71.1% accuracy on VideoMME without and with subtitles, respectively, and consistently outperforming LongVILA-7B across multiple benchmarks. Moreover, LongVILA-R1-7B supports processing up to 8,192 video frames per video, and configurable FPS settings. Notably, our MR-SP system achieves up to 2.1x speedup on long video RL training. In addition, we release our training system for public availability that supports RL training on various modalities (video, text, and audio), various models (VILA and Qwen series), and even image and video generation models. On a single A100 node (8 GPUs), it supports RL training on hour-long videos (e.g., 3,600 frames).

Junwen Xiong, Peng Zhang, Tao You et al.

Audio-visual saliency prediction can draw support from diverse modality complements, but further performance enhancement is still challenged by customized architectures as well as task-specific loss functions. In recent studies, denoising diffusion models have shown more promising in unifying task frameworks owing to their inherent ability of generalization. Following this motivation, a novel Diffusion architecture for generalized audio-visual Saliency prediction (DiffSal) is proposed in this work, which formulates the prediction problem as a conditional generative task of the saliency map by utilizing input audio and video as the conditions. Based on the spatio-temporal audio-visual features, an extra network Saliency-UNet is designed to perform multi-modal attention modulation for progressive refinement of the ground-truth saliency map from the noisy map. Extensive experiments demonstrate that the proposed DiffSal can achieve excellent performance across six challenging audio-visual benchmarks, with an average relative improvement of 6.3\% over the previous state-of-the-art results by six metrics.

Andi Han, Wei Huang, Yuan Cao et al.

The predominant success of diffusion models in generative modeling has spurred significant interest in understanding their theoretical foundations. In this work, we propose a feature learning framework aimed at analyzing and comparing the training dynamics of diffusion models with those of traditional classification models. Our theoretical analysis demonstrates that diffusion models, due to the denoising objective, are encouraged to learn more balanced and comprehensive representations of the data. In contrast, neural networks with a similar architecture trained for classification tend to prioritize learning specific patterns in the data, often focusing on easy-to-learn components. To support these theoretical insights, we conduct several experiments on both synthetic and real-world datasets, which empirically validate our findings and highlight the distinct feature learning dynamics in diffusion models compared to classification.

Bikang Pan, Qun Li, Xiaoying Tang et al.

The emergence of vision-language foundation models, such as CLIP, has revolutionized image-text representation, enabling a broad range of applications via prompt learning. Despite its promise, real-world datasets often contain noisy labels that can degrade prompt learning performance. In this paper, we demonstrate that using mean absolute error (MAE) loss in prompt learning, named PromptMAE, significantly enhances robustness against noisy labels while maintaining high accuracy. Though MAE is straightforward and recognized for its robustness, it is rarely used in noisy-label learning due to its slow convergence and poor performance outside prompt learning scenarios. To elucidate the robustness of PromptMAE, we leverage feature learning theory to show that MAE can suppress the influence of noisy samples, thereby improving the signal-to-noise ratio and enhancing overall robustness. Additionally, we introduce PromptOT, a prompt-based optimal transport data purification method to enhance the robustness further. PromptOT employs text features in vision-language models as prototypes to construct an optimal transportation matrix. This matrix effectively partitions datasets into clean and noisy subsets, allowing for the application of cross-entropy loss to the clean subset and MAE loss to the noisy subset. Our Noise-Label Prompt Learning method, named NLPrompt, offers a simple and efficient approach that leverages the expressive representations and precise alignment capabilities of vision-language models for robust prompt learning. We validate NLPrompt through extensive experiments across various noise settings, demonstrating significant performance improvements.

Yintao Zhou, Wei Huang, Zhengyu Li et al.

Parkinson's disease (PD) severity diagnosis is crucial for early detecting potential patients and adopting tailored interventions. Diagnosing PD based on facial expression is grounded in PD patients' "masked face" symptom and gains growing interest recently for its convenience and affordability. However, current facial expression-based approaches often rely on single type of expression which can lead to misdiagnosis, and ignore the class imbalance across different PD stages which degrades the prediction performance. Moreover, most existing methods focus on binary classification (i.e., PD / non-PD) rather than diagnosing the severity of PD. To address these issues, we propose a new facial expression-based method for PD severity diagnosis which integrates multiple facial expression features through attention-based feature fusion. Moreover, we mitigate the class imbalance problem via an adaptive class balancing strategy which dynamically adjusts the contribution of training samples based on their class distribution and classification difficulty. Experimental results demonstrate the promising performance of the proposed method for PD severity diagnosis, as well as the efficacy of attention-based feature fusion and adaptive class balancing.

Qiao Sun, Liujia Yang, Wei Tang et al.

While video-generation-based embodied world models have gained increasing attention, their reliance on large-scale embodied interaction data remains a key bottleneck. The scarcity, difficulty of collection, and high dimensionality of embodied data fundamentally limit the alignment granularity between language and actions and exacerbate the challenge of long-horizon video generation--hindering generative models from achieving a "GPT moment" in the embodied domain. There is a naive observation: the diversity of embodied data far exceeds the relatively small space of possible primitive motions. Based on this insight, we propose a novel paradigm for world modeling--Primitive Embodied World Models (PEWM). By restricting video generation to fixed short horizons, our approach 1) enables fine-grained alignment between linguistic concepts and visual representations of robotic actions, 2) reduces learning complexity, 3) improves data efficiency in embodied data collection, and 4) decreases inference latency. By equipping with a modular Vision-Language Model (VLM) planner and a Start-Goal heatmap Guidance mechanism (SGG), PEWM further enables flexible closed-loop control and supports compositional generalization of primitive-level policies over extended, complex tasks. Our framework leverages the spatiotemporal vision priors in video models and the semantic awareness of VLMs to bridge the gap between fine-grained physical interaction and high-level reasoning, paving the way toward scalable, interpretable, and general-purpose embodied intelligence.

Jingxian Sun, Lichao Zhang, Yufei Zha et al.

The target representation learned by convolutional neural networks plays an important role in Thermal Infrared (TIR) tracking. Currently, most of the top-performing TIR trackers are still employing representations learned by the model trained on the RGB data. However, this representation does not take into account the information in the TIR modality itself, limiting the performance of TIR tracking. To solve this problem, we propose to distill representations of the TIR modality from the RGB modality with Cross-Modal Distillation (CMD) on a large amount of unlabeled paired RGB-TIR data. We take advantage of the two-branch architecture of the baseline tracker, i.e. DiMP, for cross-modal distillation working on two components of the tracker. Specifically, we use one branch as a teacher module to distill the representation learned by the model into the other branch. Benefiting from the powerful model in the RGB modality, the cross-modal distillation can learn the TIR-specific representation for promoting TIR tracking. The proposed approach can be incorporated into different baseline trackers conveniently as a generic and independent component. Furthermore, the semantic coherence of paired RGB and TIR images is utilized as a supervised signal in the distillation loss for cross-modal knowledge transfer. In practice, three different approaches are explored to generate paired RGB-TIR patches with the same semantics for training in an unsupervised way. It is easy to extend to an even larger scale of unlabeled training data. Extensive experiments on the LSOTB-TIR dataset and PTB-TIR dataset demonstrate that our proposed cross-modal distillation method effectively learns TIR-specific target representations transferred from the RGB modality. Our tracker outperforms the baseline tracker by achieving absolute gains of 2.3% Success, 2.7% Precision, and 2.5% Normalized Precision respectively.