Abstract:We propose a novel spatial-temporal graph Mamba (STG-Mamba) for the music-guided dance video synthesis task, i.e., to translate the input music to a dance video. STG-Mamba consists of two translation mappings: music-to-skeleton translation and skeleton-to-video translation. In the music-to-skeleton translation, we introduce a novel spatial-temporal graph Mamba (STGM) block to effectively construct skeleton sequences from the input music, capturing dependencies between joints in both the spatial and temporal dimensions. For the skeleton-to-video translation, we propose a novel self-supervised regularization network to translate the generated skeletons, along with a conditional image, into a dance video. Lastly, we collect a new skeleton-to-video translation dataset from the Internet, containing 54,944 video clips. Extensive experiments demonstrate that STG-Mamba achieves significantly better results than existing methods.
Abstract:Despite the success of deep learning across various domains, it remains vulnerable to adversarial attacks. Although many existing adversarial attack methods achieve high success rates, they typically rely on $\ell_{p}$-norm perturbation constraints, which do not align with human perceptual capabilities. Consequently, researchers have shifted their focus toward generating natural, unrestricted adversarial examples (UAEs). GAN-based approaches suffer from inherent limitations, such as poor image quality due to instability and mode collapse. Meanwhile, diffusion models have been employed for UAE generation, but they still rely on iterative PGD perturbation injection, without fully leveraging their central denoising capabilities. In this paper, we introduce a novel approach for generating UAEs based on diffusion models, named ScoreAdv. This method incorporates an interpretable adversarial guidance mechanism to gradually shift the sampling distribution towards the adversarial distribution, while using an interpretable saliency map to inject the visual information of a reference image into the generated samples. Notably, our method is capable of generating an unlimited number of natural adversarial examples and can attack not only classification models but also retrieval models. We conduct extensive experiments on ImageNet and CelebA datasets, validating the performance of ScoreAdv across ten target models in both black-box and white-box settings. Our results demonstrate that ScoreAdv achieves state-of-the-art attack success rates and image quality. Furthermore, the dynamic balance between denoising and adversarial perturbation enables ScoreAdv to remain robust even under defensive measures.
Abstract:Scaling test-time compute brings substantial performance gains for large language models (LLMs). By sampling multiple answers and heuristically aggregate their answers (e.g., either through majority voting or using verifiers to rank the answers), one can achieve consistent performance gains in math domains. In this paper, we propose a new way to leverage such multiple sample set. We train a compact LLM, called Sample Set Aggregator (SSA), that takes a concatenated sequence of multiple samples and output the final answer, optimizing it for the answer accuracy with reinforcement learning. Experiments on multiple reasoning datasets show that SSA outperforms other test-time scaling methods such as reward model-based re-ranking. Our approach also shows a promising generalization ability, across sample set sizes, base model families and scales, and tasks. By separating LLMs to generate answers and LLMs to analyze and aggregate sampled answers, our approach can work with the outputs from premier black box models easily and efficiently.
Abstract:Large multimodal models (LMMs) suffer significant computational challenges due to the high cost of Large Language Models (LLMs) and the quadratic complexity of processing long vision token sequences. In this paper, we explore the spatial redundancy among vision tokens and shorten the length of vision token sequences for inference acceleration. Specifically, we propose a Spatial Token Fusion (STF) method to learn compact vision tokens for short vision token sequence, where spatial-adjacent tokens are fused into one. Meanwhile, weight-frozen vision encoder can not well adapt to the demand of extensive downstream vision-language tasks. To this end, we further introduce a Multi-Block Token Fusion (MBTF) module to supplement multi-granularity features for the reduced token sequence. Overall, we combine STF and MBTF module to balance token reduction and information preservation, thereby improving inference efficiency without sacrificing multimodal reasoning capabilities. Experimental results demonstrate that our method based on LLaVA-1.5 achieves comparable or even superior performance to the baseline on 8 popular vision-language benchmarks with only $25\%$ vision tokens of baseline. The source code and trained weights are available at https://github.com/visresearch/LLaVA-STF.
Abstract:Vision transformers have been widely explored in various vision tasks. Due to heavy computational cost, much interest has aroused for compressing vision transformer dynamically in the aspect of tokens. Current methods mainly pay attention to token pruning or merging to reduce token numbers, in which tokens are compressed exclusively, causing great information loss and therefore post-training is inevitably required to recover the performance. In this paper, we rethink token reduction and unify the process as an explicit form of token matrix transformation, in which all existing methods are constructing special forms of matrices within the framework. Furthermore, we propose a many-to-many Token Transforming framework that serves as a generalization of all existing methods and reserves the most information, even enabling training-free acceleration. We conduct extensive experiments to validate our framework. Specifically, we reduce 40% FLOPs and accelerate DeiT-S by $\times$1.5 with marginal 0.1% accuracy drop. Furthermore, we extend the method to dense prediction tasks including segmentation, object detection, depth estimation, and language model generation. Results demonstrate that the proposed method consistently achieves substantial improvements, offering a better computation-performance trade-off, impressive budget reduction and inference acceleration.
Abstract:Modern neural speech models benefit from having longer context, and many approaches have been proposed to increase the maximum context a model can use. However, few have attempted to measure how much context these models actually use, i.e., the effective context. Here, we propose two approaches to measuring the effective context, and use them to analyze different speech Transformers. For supervised models, we find that the effective context correlates well with the nature of the task, with fundamental frequency tracking, phone classification, and word classification requiring increasing amounts of effective context. For self-supervised models, we find that effective context increases mainly in the early layers, and remains relatively short -- similar to the supervised phone model. Given that these models do not use a long context during prediction, we show that HuBERT can be run in streaming mode without modification to the architecture and without further fine-tuning.
Abstract:Large language models (LLMs) have shown remarkable promise but remain challenging to continually improve through traditional finetuning, particularly when integrating capabilities from other specialized LLMs. Popular methods like ensemble and weight merging require substantial memory and struggle to adapt to changing data environments. Recent efforts have transferred knowledge from multiple LLMs into a single target model; however, they suffer from interference and degraded performance among tasks, largely due to limited flexibility in candidate selection and training pipelines. To address these issues, we propose a framework that adaptively selects and aggregates knowledge from diverse LLMs to build a single, stronger model, avoiding the high memory overhead of ensemble and inflexible weight merging. Specifically, we design an adaptive selection network that identifies the most relevant source LLMs based on their scores, thereby reducing knowledge interference. We further propose a dynamic weighted fusion strategy that accounts for the inherent strengths of candidate LLMs, along with a feedback-driven loss function that prevents the selector from converging on a single subset of sources. Experimental results demonstrate that our method can enable a more stable and scalable knowledge aggregation process while reducing knowledge interference by up to 50% compared to existing approaches. Code is avaliable at https://github.com/ZLKong/LLM_Integration
Abstract:Conventional frame-based cameras often struggle with stereo depth estimation in rapidly changing scenes. In contrast, bio-inspired spike cameras emit asynchronous events at microsecond-level resolution, providing an alternative sensing modality. However, existing methods lack specialized stereo algorithms and benchmarks tailored to the spike data. To address this gap, we propose SpikeStereoNet, a brain-inspired framework and the first to estimate stereo depth directly from raw spike streams. The model fuses raw spike streams from two viewpoints and iteratively refines depth estimation through a recurrent spiking neural network (RSNN) update module. To benchmark our approach, we introduce a large-scale synthetic spike stream dataset and a real-world stereo spike dataset with dense depth annotations. SpikeStereoNet outperforms existing methods on both datasets by leveraging spike streams' ability to capture subtle edges and intensity shifts in challenging regions such as textureless surfaces and extreme lighting conditions. Furthermore, our framework exhibits strong data efficiency, maintaining high accuracy even with substantially reduced training data. The source code and datasets will be publicly available.
Abstract:In Transformer architectures, tokens\textemdash discrete units derived from raw data\textemdash are formed by segmenting inputs into fixed-length chunks. Each token is then mapped to an embedding, enabling parallel attention computations while preserving the input's essential information. Due to the quadratic computational complexity of transformer self-attention mechanisms, token reduction has primarily been used as an efficiency strategy. This is especially true in single vision and language domains, where it helps balance computational costs, memory usage, and inference latency. Despite these advances, this paper argues that token reduction should transcend its traditional efficiency-oriented role in the era of large generative models. Instead, we position it as a fundamental principle in generative modeling, critically influencing both model architecture and broader applications. Specifically, we contend that across vision, language, and multimodal systems, token reduction can: (i) facilitate deeper multimodal integration and alignment, (ii) mitigate "overthinking" and hallucinations, (iii) maintain coherence over long inputs, and (iv) enhance training stability, etc. We reframe token reduction as more than an efficiency measure. By doing so, we outline promising future directions, including algorithm design, reinforcement learning-guided token reduction, token optimization for in-context learning, and broader ML and scientific domains. We highlight its potential to drive new model architectures and learning strategies that improve robustness, increase interpretability, and better align with the objectives of generative modeling.
Abstract:To address the challenge of complex pathological feature extraction in automated cardiac MRI segmentation, this study proposes an innovative dual-encoder architecture named SAMba-UNet. The framework achieves cross-modal feature collaborative learning by integrating the vision foundation model SAM2, the state-space model Mamba, and the classical UNet. To mitigate domain discrepancies between medical and natural images, a Dynamic Feature Fusion Refiner is designed, which enhances small lesion feature extraction through multi-scale pooling and a dual-path calibration mechanism across channel and spatial dimensions. Furthermore, a Heterogeneous Omni-Attention Convergence Module (HOACM) is introduced, combining global contextual attention with branch-selective emphasis mechanisms to effectively fuse SAM2's local positional semantics and Mamba's long-range dependency modeling capabilities. Experiments on the ACDC cardiac MRI dataset demonstrate that the proposed model achieves a Dice coefficient of 0.9103 and an HD95 boundary error of 1.0859 mm, significantly outperforming existing methods, particularly in boundary localization for complex pathological structures such as right ventricular anomalies. This work provides an efficient and reliable solution for automated cardiac disease diagnosis, and the code will be open-sourced.