Yansong Tang

Haipeng Luo, Qingfeng Sun, Can Xu et al. · microsoft-research

Large language models (LLMs), such as GPT-4, have shown remarkable performance in natural language processing (NLP) tasks, including challenging mathematical reasoning. However, most existing open-source models are only pre-trained on large-scale internet data and without math-related optimization. In this paper, we present WizardMath, which enhances the mathematical CoT reasoning abilities of LLMs without using external python tools, by applying our proposed Reinforcement Learning from Evol-Instruct Feedback (RLEIF) method to the domain of math. Through extensive experiments on two mathematical reasoning benchmarks, namely GSM8k and MATH, we reveal the extraordinary capabilities of our model. Remarkably, WizardMath-Mistral 7B surpasses top-tier open-source LLMs by a substantial margin with higher data efficiency. Furthermore, WizardMath 70B even outperforms GPT-3.5-Turbo, Claude 2, Gemini Pro and GPT-4-early-version. Additionally, our preliminary exploration highlights the pivotal role of instruction evolution and process supervision in achieving exceptional math performance. For more details refer to https://github.com/nlpxucan/WizardLM

Yongming Rao, Wenliang Zhao, Yansong Tang et al. · tsinghua

Recent progress in vision Transformers exhibits great success in various tasks driven by the new spatial modeling mechanism based on dot-product self-attention. In this paper, we show that the key ingredients behind the vision Transformers, namely input-adaptive, long-range and high-order spatial interactions, can also be efficiently implemented with a convolution-based framework. We present the Recursive Gated Convolution ($\textit{g}^\textit{n}$Conv) that performs high-order spatial interactions with gated convolutions and recursive designs. The new operation is highly flexible and customizable, which is compatible with various variants of convolution and extends the two-order interactions in self-attention to arbitrary orders without introducing significant extra computation. $\textit{g}^\textit{n}$Conv can serve as a plug-and-play module to improve various vision Transformers and convolution-based models. Based on the operation, we construct a new family of generic vision backbones named HorNet. Extensive experiments on ImageNet classification, COCO object detection and ADE20K semantic segmentation show HorNet outperform Swin Transformers and ConvNeXt by a significant margin with similar overall architecture and training configurations. HorNet also shows favorable scalability to more training data and larger model sizes. Apart from the effectiveness in visual encoders, we also show $\textit{g}^\textit{n}$Conv can be applied to task-specific decoders and consistently improve dense prediction performance with less computation. Our results demonstrate that $\textit{g}^\textit{n}$Conv can be a new basic module for visual modeling that effectively combines the merits of both vision Transformers and CNNs. Code is available at https://github.com/raoyongming/HorNet

Wanhua Li, Xiaoke Huang, Zheng Zhu et al. · tsinghua

This paper presents a language-powered paradigm for ordinal regression. Existing methods usually treat each rank as a category and employ a set of weights to learn these concepts. These methods are easy to overfit and usually attain unsatisfactory performance as the learned concepts are mainly derived from the training set. Recent large pre-trained vision-language models like CLIP have shown impressive performance on various visual tasks. In this paper, we propose to learn the rank concepts from the rich semantic CLIP latent space. Specifically, we reformulate this task as an image-language matching problem with a contrastive objective, which regards labels as text and obtains a language prototype from a text encoder for each rank. While prompt engineering for CLIP is extremely time-consuming, we propose OrdinalCLIP, a differentiable prompting method for adapting CLIP for ordinal regression. OrdinalCLIP consists of learnable context tokens and learnable rank embeddings; The learnable rank embeddings are constructed by explicitly modeling numerical continuity, resulting in well-ordered, compact language prototypes in the CLIP space. Once learned, we can only save the language prototypes and discard the huge language model, resulting in zero additional computational overhead compared with the linear head counterpart. Experimental results show that our paradigm achieves competitive performance in general ordinal regression tasks, and gains improvements in few-shot and distribution shift settings for age estimation. The code is available at https://github.com/xk-huang/OrdinalCLIP.

Benlin Liu, Yuhao Dong, Yiqin Wang et al. · deepmind, tsinghua

Multimodal language models (MLLMs) are increasingly being applied in real-world environments, necessitating their ability to interpret 3D spaces and comprehend temporal dynamics. Current methods often rely on specialized architectural designs or task-specific fine-tuning to achieve this. We introduce Coarse Correspondences, a simple lightweight method that enhances MLLMs' spatial-temporal reasoning with 2D images as input, without modifying the architecture or requiring task-specific fine-tuning. Our method uses a lightweight tracking model to identify primary object correspondences between frames in a video or across different image viewpoints, and then conveys this information to MLLMs through visual prompting. We demonstrate that this simple training-free approach brings substantial gains to GPT4-V/O consistently on four benchmarks that require spatial-temporal reasoning, including +20.5\% improvement on ScanQA, +9.7\% on OpenEQA's episodic memory subset, +6.0\% on the long-form video benchmark EgoSchema, and +11\% on the R2R navigation benchmark. Additionally, we show that Coarse Correspondences can also enhance open-source MLLMs' spatial reasoning (by +6.9\% on ScanQA) when applied in both training and inference and that the improvement can generalize to unseen datasets such as SQA3D (+3.1\%). Taken together, we show that Coarse Correspondences effectively and efficiently boosts models' performance on downstream tasks requiring spatial-temporal reasoning.

Yansong Tang, Xingyu Liu, Xumin Yu et al.

Rapid progress and superior performance have been achieved for skeleton-based action recognition recently. In this article, we investigate this problem under a cross-dataset setting, which is a new, pragmatic, and challenging task in real-world scenarios. Following the unsupervised domain adaptation (UDA) paradigm, the action labels are only available on a source dataset, but unavailable on a target dataset in the training stage. Different from the conventional adversarial learning-based approaches for UDA, we utilize a self-supervision scheme to reduce the domain shift between two skeleton-based action datasets. Our inspiration is drawn from Cubism, an art genre from the early 20th century, which breaks and reassembles the objects to convey a greater context. By segmenting and permuting temporal segments or human body parts, we design two self-supervised learning classification tasks to explore the temporal and spatial dependency of a skeleton-based action and improve the generalization ability of the model. We conduct experiments on six datasets for skeleton-based action recognition, including three large-scale datasets (NTU RGB+D, PKU-MMD, and Kinetics) where new cross-dataset settings and benchmarks are established. Extensive results demonstrate that our method outperforms state-of-the-art approaches. The source codes of our model and all the compared methods are available at https://github.com/shanice-l/st-cubism.

Yong Liu, Ran Yu, Jiahao Wang et al.

This paper studies semi-supervised video object segmentation through boosting intra-frame interaction. Recent memory network-based methods focus on exploiting inter-frame temporal reference while paying little attention to intra-frame spatial dependency. Specifically, these segmentation model tends to be susceptible to interference from unrelated nontarget objects in a certain frame. To this end, we propose Global Spectral Filter Memory network (GSFM), which improves intra-frame interaction through learning long-term spatial dependencies in the spectral domain. The key components of GSFM is 2D (inverse) discrete Fourier transform for spatial information mixing. Besides, we empirically find low frequency feature should be enhanced in encoder (backbone) while high frequency for decoder (segmentation head). We attribute this to semantic information extracting role for encoder and fine-grained details highlighting role for decoder. Thus, Low (High) Frequency Module is proposed to fit this circumstance. Extensive experiments on the popular DAVIS and YouTube-VOS benchmarks demonstrate that GSFM noticeably outperforms the baseline method and achieves state-of-the-art performance. Besides, extensive analysis shows that the proposed modules are reasonable and of great generalization ability. Our source code is available at https://github.com/workforai/GSFM.

Ronghui Li, Junfan Zhao, Yachao Zhang et al. · tsinghua

Generating full-body and multi-genre dance sequences from given music is a challenging task, due to the limitations of existing datasets and the inherent complexity of the fine-grained hand motion and dance genres. To address these problems, we propose FineDance, which contains 14.6 hours of music-dance paired data, with fine-grained hand motions, fine-grained genres (22 dance genres), and accurate posture. To the best of our knowledge, FineDance is the largest music-dance paired dataset with the most dance genres. Additionally, to address monotonous and unnatural hand movements existing in previous methods, we propose a full-body dance generation network, which utilizes the diverse generation capabilities of the diffusion model to solve monotonous problems, and use expert nets to solve unreal problems. To further enhance the genre-matching and long-term stability of generated dances, we propose a Genre&Coherent aware Retrieval Module. Besides, we propose a novel metric named Genre Matching Score to evaluate the genre-matching degree between dance and music. Quantitative and qualitative experiments demonstrate the quality of FineDance, and the state-of-the-art performance of FineNet. The FineDance Dataset and more qualitative samples can be found at our website.

Yansong Tang, Jinpeng Liu, Aoyang Liu et al. · tsinghua

With the continuously thriving popularity around the world, fitness activity analytic has become an emerging research topic in computer vision. While a variety of new tasks and algorithms have been proposed recently, there are growing hunger for data resources involved in high-quality data, fine-grained labels, and diverse environments. In this paper, we present FLAG3D, a large-scale 3D fitness activity dataset with language instruction containing 180K sequences of 60 categories. FLAG3D features the following three aspects: 1) accurate and dense 3D human pose captured from advanced MoCap system to handle the complex activity and large movement, 2) detailed and professional language instruction to describe how to perform a specific activity, 3) versatile video resources from a high-tech MoCap system, rendering software, and cost-effective smartphones in natural environments. Extensive experiments and in-depth analysis show that FLAG3D contributes great research value for various challenges, such as cross-domain human action recognition, dynamic human mesh recovery, and language-guided human action generation. Our dataset and source code are publicly available at https://andytang15.github.io/FLAG3D.

Haipeng Luo, Qingfeng Sun, Can Xu et al. · microsoft-research

Assessing the effectiveness of large language models (LLMs) presents substantial challenges. The method of conducting human-annotated battles in an online Chatbot Arena is a highly effective evaluative technique. However, this approach is limited by the costs and time required for human annotation. In this paper, we introduce Arena Learning, an innovative offline strategy designed to simulate these arena battles using AI-driven annotations to evaluate battle outcomes, thus facilitating the continuous improvement of the target model through both supervised fine-tuning and reinforcement learning. Arena Learning comprises two key elements. First, it ensures precise evaluations and maintains consistency between offline simulations and online competitions via WizardArena, a pipeline developed to accurately predict the Elo rankings of various models using a meticulously designed offline test set. Our results demonstrate that WizardArena's predictions closely align with those from the online Arena. Second, it involves the continuous improvement of training data based on the battle results and the refined model. We establish a data flywheel to iteratively update the training data by highlighting the weaknesses of the target model based on its battle results, enabling it to learn from the strengths of multiple different models. We apply Arena Learning to train our target model, WizardLM-$β$, and demonstrate significant performance enhancements across various metrics. This fully automated training and evaluation pipeline sets the stage for continuous advancements in various LLMs via post-training. Notably, Arena Learning plays a pivotal role in the success of WizardLM-2, and this paper serves both as an exploration of its efficacy and a foundational study for future discussions related to WizardLM-2 and its derivatives.

Yinan Liang, Ziwei Wang, Xiuwei Xu et al.

Due to the high price and heavy energy consumption of GPUs, deploying deep models on IoT devices such as microcontrollers makes significant contributions for ecological AI. Conventional methods successfully enable convolutional neural network inference of high resolution images on microcontrollers, while the framework for vision transformers that achieve the state-of-the-art performance in many vision applications still remains unexplored. In this paper, we propose a hardware-algorithm co-optimizations method called MCUFormer to deploy vision transformers on microcontrollers with extremely limited memory, where we jointly design transformer architecture and construct the inference operator library to fit the memory resource constraint. More specifically, we generalize the one-shot network architecture search (NAS) to discover the optimal architecture with highest task performance given the memory budget from the microcontrollers, where we enlarge the existing search space of vision transformers by considering the low-rank decomposition dimensions and patch resolution for memory reduction. For the construction of the inference operator library of vision transformers, we schedule the memory buffer during inference through operator integration, patch embedding decomposition, and token overwriting, allowing the memory buffer to be fully utilized to adapt to the forward pass of the vision transformer. Experimental results demonstrate that our MCUFormer achieves 73.62\% top-1 accuracy on ImageNet for image classification with 320KB memory on STM32F746 microcontroller. Code is available at https://github.com/liangyn22/MCUFormer.

Xiaoke Huang, Yiji Cheng, Yansong Tang et al. · tsinghua

Efficiently digitizing high-fidelity animatable human avatars from videos is a challenging and active research topic. Recent volume rendering-based neural representations open a new way for human digitization with their friendly usability and photo-realistic reconstruction quality. However, they are inefficient for long optimization times and slow inference speed; their implicit nature results in entangled geometry, materials, and dynamics of humans, which are hard to edit afterward. Such drawbacks prevent their direct applicability to downstream applications, especially the prominent rasterization-based graphic ones. We present EMA, a method that Efficiently learns Meshy neural fields to reconstruct animatable human Avatars. It jointly optimizes explicit triangular canonical mesh, spatial-varying material, and motion dynamics, via inverse rendering in an end-to-end fashion. Each above component is derived from separate neural fields, relaxing the requirement of a template, or rigging. The mesh representation is highly compatible with the efficient rasterization-based renderer, thus our method only takes about an hour of training and can render in real-time. Moreover, only minutes of optimization is enough for plausible reconstruction results. The disentanglement of meshes enables direct downstream applications. Extensive experiments illustrate the very competitive performance and significant speed boost against previous methods. We also showcase applications including novel pose synthesis, material editing, and relighting. The project page: https://xk-huang.github.io/ema/.

Kunyang Han, Yong Liu, Jun Hao Liew et al.

Recent advancements in pre-trained vision-language models, such as CLIP, have enabled the segmentation of arbitrary concepts solely from textual inputs, a process commonly referred to as open-vocabulary semantic segmentation (OVS). However, existing OVS techniques confront a fundamental challenge: the trained classifier tends to overfit on the base classes observed during training, resulting in suboptimal generalization performance to unseen classes. To mitigate this issue, recent studies have proposed the use of an additional frozen pre-trained CLIP for classification. Nonetheless, this approach incurs heavy computational overheads as the CLIP vision encoder must be repeatedly forward-passed for each mask, rendering it impractical for real-world applications. To address this challenge, our objective is to develop a fast OVS model that can perform comparably or better without the extra computational burden of the CLIP image encoder during inference. To this end, we propose a core idea of preserving the generalizable representation when fine-tuning on known classes. Specifically, we introduce a text diversification strategy that generates a set of synonyms for each training category, which prevents the learned representation from collapsing onto specific known category names. Additionally, we employ a text-guided knowledge distillation method to preserve the generalizable knowledge of CLIP. Extensive experiments demonstrate that our proposed model achieves robust generalization performance across various datasets. Furthermore, we perform a preliminary exploration of open-vocabulary video segmentation and present a benchmark that can facilitate future open-vocabulary research in the video domain.

Kejie Li, Yansong Tang, Victor Adrian Prisacariu et al.

Dense 3D reconstruction from a stream of depth images is the key to many mixed reality and robotic applications. Although methods based on Truncated Signed Distance Function (TSDF) Fusion have advanced the field over the years, the TSDF volume representation is confronted with striking a balance between the robustness to noisy measurements and maintaining the level of detail. We present Bi-level Neural Volume Fusion (BNV-Fusion), which leverages recent advances in neural implicit representations and neural rendering for dense 3D reconstruction. In order to incrementally integrate new depth maps into a global neural implicit representation, we propose a novel bi-level fusion strategy that considers both efficiency and reconstruction quality by design. We evaluate the proposed method on multiple datasets quantitatively and qualitatively, demonstrating a significant improvement over existing methods.

Zhao Yang, Jiaqi Wang, Yansong Tang et al.

Referring image segmentation segments an image from a language expression. With the aim of producing high-quality masks, existing methods often adopt iterative learning approaches that rely on RNNs or stacked attention layers to refine vision-language features. Despite their complexity, RNN-based methods are subject to specific encoder choices, while attention-based methods offer limited gains. In this work, we introduce a simple yet effective alternative for progressively learning discriminative multi-modal features. The core idea of our approach is to leverage a continuously updated query as the representation of the target object and at each iteration, strengthen multi-modal features strongly correlated to the query while weakening less related ones. As the query is initialized by language features and successively updated by object features, our algorithm gradually shifts from being localization-centric to segmentation-centric. This strategy enables the incremental recovery of missing object parts and/or removal of extraneous parts through iteration. Compared to its counterparts, our method is more versatile$\unicode{x2014}$it can be plugged into prior arts straightforwardly and consistently bring improvements. Experimental results on the challenging datasets of RefCOCO, RefCOCO+, and G-Ref demonstrate its advantage with respect to the state-of-the-art methods.

Yue Wu, Changyuan Wang, Zixuan Wang et al.

Conventional Post-Training Quantization (PTQ) methods struggle with 4-bit Omni-modal Large Language Models (OLLMs) due to the extreme distribution heterogeneity and disparate outlier patterns across modalities. To address this, we propose MorphoQuant, a modality-aware PTQ framework engineered to preserve cross-modal morphology and mitigate outlier loss. Specifically, we introduce Distribution-Aware Bias Compensation (DABC), which selectively absorbs long-tailed outliers into channel-wise biases. This mechanism safeguards outlier magnitudes while maintaining high-precision discretization for dense inliers, thereby preserving accurate discretization across diverse modal distribution. Complementing this, we propose Morphology-Directed Quantization Function Optimization (MDQFO) to co-optimize the quantization grid with the bias mask, ensuring fine-grained alignment across modalities. Extensive evaluations on Qwen2.5-Omni across benchmarks like MMMU and Video-MME demonstrate our approach's superiority. Notably, our W4A4 model achieves 76.63% on ScienceQA, significantly outperforming SOTA W4A4 methods and surprisingly surpassing the W4A16 baseline, which fully demonstrates the exceptional accuracy-efficiency trade-off of our framework.

Runzhong Zhang, Suchen Wang, Yueqi Duan et al.

In this paper, we propose an HOI-aware adaptive network named AdaAct for weakly-supervised action segmentation. Most existing methods learn a fixed network to predict the action of each frame with the neighboring frames. However, this would result in ambiguity when estimating similar actions, such as pouring juice and pouring coffee. To address this, we aim to exploit temporally global but spatially local human-object interactions (HOI) as video-level prior knowledge for action segmentation. The long-term HOI sequence provides crucial contextual information to distinguish ambiguous actions, where our network dynamically adapts to the given HOI sequence at test time. More specifically, we first design a video HOI encoder that extracts, selects, and integrates the most representative HOI throughout the video. Then, we propose a two-branch HyperNetwork to learn an adaptive temporal encoder, which automatically adjusts the parameters based on the HOI information of various videos on the fly. Extensive experiments on two widely-used datasets including Breakfast and 50Salads demonstrate the effectiveness of our method under different evaluation metrics.

Xin Dong, Wenjia Geng, Wenfeng Deng et al.

Video Moment Retrieval (MR) and Highlight Detection (HD) are crucial tasks in video analysis that aim to localize specific moments and estimate clip-wise relevance based on a given text query. Recent approaches treat them as similar video grounding tasks and use the same architecture to solve them. These tasks require both fine-grained comprehension at the image level and high-level temporal understanding across the entire video. Existing approaches have primarily focused on temporal modeling using frame-level features, often neglecting the rich visual information related to the text query within individual frames. This oversight leads to inaccurate grounding results. To address this limitation, we propose a Comprehensive Spatial-Temporal Representation Learning Framework (CoSTL), which captures both fine-grained image-level information and temporal dynamics. Specifically, CoSTL incorporates a text-driven progressive fine-grained image encoder, performing a two-step text-driven knowledge extraction process to learn fine-grained spatial representations. Furthermore, a multi-scale temporal perception module captures comprehensive spatial-temporal representations, enhancing the model's ability to process temporal dynamics. We demonstrate state-of-the-art performance on four public benchmarks: QVHighlights, Charades-STA, TACoS, and TVSum.

Rui Yang, Hailong Ma, Jie Wu et al.

The vanilla self-attention mechanism inherently relies on pre-defined and steadfast computational dimensions. Such inflexibility restricts it from possessing context-oriented generalization that can bring more contextual cues and global representations. To mitigate this issue, we propose a Scalable Self-Attention (SSA) mechanism that leverages two scaling factors to release dimensions of query, key, and value matrices while unbinding them with the input. This scalability fetches context-oriented generalization and enhances object sensitivity, which pushes the whole network into a more effective trade-off state between accuracy and cost. Furthermore, we propose an Interactive Window-based Self-Attention (IWSA), which establishes interaction between non-overlapping regions by re-merging independent value tokens and aggregating spatial information from adjacent windows. By stacking the SSA and IWSA alternately, the Scalable Vision Transformer (ScalableViT) achieves state-of-the-art performance in general-purpose vision tasks. For example, ScalableViT-S outperforms Twins-SVT-S by 1.4% and Swin-T by 1.8% on ImageNet-1K classification.

Deyi Zhu, Yuji Wang, Yong Liu et al.

Traditional visual object tracking (VOT) methods typically rely on task-specific supervised training, limiting their generalization to unseen objects and challenging scenarios with distractors, occlusion, and nonlinear motion. Recent vision foundation models, exemplified by SAM 2, learn strong video understanding priors from large-scale pretraining and offer a promising foundation for building more robust and generalizable trackers. However, directly applying SAM 2 to VOT remains suboptimal, as it does not explicitly model target motion dynamics or enforce geometric and semantic consistency across frames, both of which are essential for reliable tracking. To address this issue, we propose SAMOSA, a new tracking framework that adapts SAM 2 to complex VOT scenarios by explicitly leveraging motion, geometry, and semantic cues. Specifically, we introduce a lightweight nonlinear motion predictor to model target dynamics and guide mask selection as well as memory filtering. We further exploit semantic cues to detect target shifts and recover from tracking failures, while geometric cues are incorporated as structural constraints to improve tracking stability. In this way, SAMOSA bridges the gap between the implicit video understanding prior of SAM 2 and explicit tracking-oriented modeling. Extensive experiments show that SAMOSA consistently outperforms state-of-the-art SAM 2--based approaches on general benchmarks, demonstrates stronger generalization than supervised VOT methods, and achieves substantial gains on anti-UAV datasets, which typify complex nonlinear motion scenarios. Our code is available at https://github.com/DurYi/SAMOSA.

Guangyi Chen, Xiao Liu, Guangrun Wang et al.

Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-Adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-Adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method.

Hongbo Liu, Mingda Wu, Kun Yuan et al.

Video quality assessment (VQA) has attracted growing attention in recent years. While the great expense of annotating large-scale VQA datasets has become the main obstacle for current deep-learning methods. To surmount the constraint of insufficient training data, in this paper, we first consider the complete range of video distribution diversity (\ie content, distortion, motion) and employ diverse pretrained models (\eg architecture, pretext task, pre-training dataset) to benefit quality representation. An Adaptive Diverse Quality-aware feature Acquisition (Ada-DQA) framework is proposed to capture desired quality-related features generated by these frozen pretrained models. By leveraging the Quality-aware Acquisition Module (QAM), the framework is able to extract more essential and relevant features to represent quality. Finally, the learned quality representation is utilized as supplementary supervisory information, along with the supervision of the labeled quality score, to guide the training of a relatively lightweight VQA model in a knowledge distillation manner, which largely reduces the computational cost during inference. Experimental results on three mainstream no-reference VQA benchmarks clearly show the superior performance of Ada-DQA in comparison with current state-of-the-art approaches without using extra training data of VQA.

Siao Tang, Xin Wang, Hong Chen et al.

High computational overhead is a troublesome problem for diffusion models. Recent studies have leveraged post-training quantization (PTQ) to compress diffusion models. However, most of them only focus on unconditional models, leaving the quantization of widely-used pretrained text-to-image models, e.g., Stable Diffusion, largely unexplored. In this paper, we propose a novel post-training quantization method PCR (Progressive Calibration and Relaxing) for text-to-image diffusion models, which consists of a progressive calibration strategy that considers the accumulated quantization error across timesteps, and an activation relaxing strategy that improves the performance with negligible cost. Additionally, we demonstrate the previous metrics for text-to-image diffusion model quantization are not accurate due to the distribution gap. To tackle the problem, we propose a novel QDiffBench benchmark, which utilizes data in the same domain for more accurate evaluation. Besides, QDiffBench also considers the generalization performance of the quantized model outside the calibration dataset. Extensive experiments on Stable Diffusion and Stable Diffusion XL demonstrate the superiority of our method and benchmark. Moreover, we are the first to achieve quantization for Stable Diffusion XL while maintaining the performance.

Zhiheng Li, Wenjia Geng, Muheng Li et al.

In this paper, we propose Skip-Plan, a condensed action space learning method for procedure planning in instructional videos. Current procedure planning methods all stick to the state-action pair prediction at every timestep and generate actions adjacently. Although it coincides with human intuition, such a methodology consistently struggles with high-dimensional state supervision and error accumulation on action sequences. In this work, we abstract the procedure planning problem as a mathematical chain model. By skipping uncertain nodes and edges in action chains, we transfer long and complex sequence functions into short but reliable ones in two ways. First, we skip all the intermediate state supervision and only focus on action predictions. Second, we decompose relatively long chains into multiple short sub-chains by skipping unreliable intermediate actions. By this means, our model explores all sorts of reliable sub-relations within an action sequence in the condensed action space. Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask and COIN benchmarks for procedure planning.

Bowen Zhang, Yiji Cheng, Chunyu Wang et al.

We present RodinHD, which can generate high-fidelity 3D avatars from a portrait image. Existing methods fail to capture intricate details such as hairstyles which we tackle in this paper. We first identify an overlooked problem of catastrophic forgetting that arises when fitting triplanes sequentially on many avatars, caused by the MLP decoder sharing scheme. To overcome this issue, we raise a novel data scheduling strategy and a weight consolidation regularization term, which improves the decoder's capability of rendering sharper details. Additionally, we optimize the guiding effect of the portrait image by computing a finer-grained hierarchical representation that captures rich 2D texture cues, and injecting them to the 3D diffusion model at multiple layers via cross-attention. When trained on 46K avatars with a noise schedule optimized for triplanes, the resulting model can generate 3D avatars with notably better details than previous methods and can generalize to in-the-wild portrait input.

Xin Dong, Yunzhi Teng, Wenfeng Deng et al.

In this work, we focus on zero-shot 3D style transfer that can generate multi-view consistent stylized views of the 3D scene given an arbitrary style image. We primarily tackle the issue of data scarcity in 3D style transfer, which arises when each model is trained on only a single scene, thereby limiting the number of available content images. This scarcity significantly hampers stylization performance, as model optimization relies on a sufficient number of content-style image pairs to provide supervisory signals. Our core idea is to integrate a decoder pre-trained on large-scale 2D image datasets into the 3D style transfer pipeline, thereby leveraging the prior knowledge encoded in the decoder from learning over numerous content-style image pairs. Our method combines feature Gaussian splatting and deferred stylization, enabling high-quality stylization with the data-sufficient decoder network while ensuring view consistency by unifying view-dependent operations into a view-invariant process. Experiments demonstrate that our Data-Sufficient StyleGaussian (DS-StyleGaussian) model outperforms existing zero-shot 3D style transfer methods in terms of visual quality across various datasets. This work also suggests that 2D pre-training can serve as a strong enhancement for 3D tasks, bridging the data gap between 2D and 3D.

Xin Dong, Weijian Deng, Lihan Zhang et al.

This work addresses the problem of recovering complete, simulatable object geometry from reconstructed real-world scenes, enabling physics-based interaction with objects embedded in the scene. While modern multi-view reconstruction methods can produce visually accurate environments, objects are often incomplete due to occlusions and limited observations, making them unsuitable for physics simulation. To address this limitation, we propose SAM3D-Phys, a framework that integrates scene reconstruction with generative 3D priors of SAM3D to recover physically simulatable objects. Our approach first reconstructs the scene from multi-view images to obtain scene geometry and partial observations of objects. We then leverage SAM3D to infer complete object geometry from these partial observations. To ensure that the recovered objects remain consistent with the reconstructed scene, we restore scene-consistent object states through two complementary strategies: a physics-constrained spatial optimization algorithm that iteratively aligns the recovered object to its original location, and a mask-guided appearance distillation module that refines texture fidelity based on the observed images. By recovering complete object geometry and restoring its pose and appearance within the scene, SAM3D-Phys produces clean object representations suitable for physics-based simulation, enabling simultaneous and physically consistent interactive simulation of multiple objects within a reconstructed scene. Project page: https://chnxindong.github.io/sam3d-phys/

Shilin Ma, Chubin Zhang, Changyuan Wang et al.

Real-time inference of vision-language-action (VLA) models is essential for robotic control. While visual token pruning has shown strong potential for accelerating inference, most existing methods mainly base pruning decisions on shallow-layer cues and risk discarding visual information required by deep layers. To address this issue, we propose SAFE-Pruner, a plug-and-play pruning framework that incorporates attention cues of future layers into pruning decisions. Specifically, we identify semantic attention consistency, the tendency that VLA models concentrate their attention probability mass on the same semantic entity across execution steps. Based on this observation, we design a forward-looking strategy to forecast the token saliency in deep layers, which prevents the premature removal of critical tokens and leads to more stable acceleration. We further introduce an adaptive subtask division strategy to detect abrupt attention shifts, thereby improving forecasting accuracy and pruning reliability. Extensive experiments in simulation and real-world settings demonstrate that our method achieves up to 1.89x speedup with a minimal degradation in success rate of less than 1.7%, while outperforming state-of-the-art methods by up to 1.9%.

Yiji Cheng, Fei Yin, Xiaoke Huang et al.

Text-to-3D is an emerging task that allows users to create 3D content with infinite possibilities. Existing works tackle the problem by optimizing a 3D representation with guidance from pre-trained diffusion models. An apparent drawback is that they need to optimize from scratch for each prompt, which is computationally expensive and often yields poor visual fidelity. In this paper, we propose DreamPortrait, which aims to generate text-guided 3D-aware portraits in a single-forward pass for efficiency. To achieve this, we extend Score Distillation Sampling from datapoint to distribution formulation, which injects semantic prior into a 3D distribution. However, the direct extension will lead to the mode collapse problem since the objective only pursues semantic alignment. Hence, we propose to optimize a distribution with hierarchical condition adapters and GAN loss regularization. For better 3D modeling, we further design a 3D-aware gated cross-attention mechanism to explicitly let the model perceive the correspondence between the text and the 3D-aware space. These elaborated designs enable our model to generate portraits with robust multi-view semantic consistency, eliminating the need for optimization-based methods. Extensive experiments demonstrate our model's highly competitive performance and significant speed boost against existing methods.

Siao Tang, Xin Wang, Hong Chen et al.

Diffusion models have recently shown remarkable generation ability, achieving state-of-the-art performance in many tasks. However, the high computational cost is still a troubling problem for diffusion models. To tackle this problem, we propose to automatically remove the structural redundancy in diffusion models with our proposed Diffusion Distillation-based Block-wise Neural Architecture Search (DiffNAS). Specifically, given a larger pretrained teacher, we leverage DiffNAS to search for the smallest architecture which can achieve on-par or even better performance than the teacher. Considering current diffusion models are based on UNet which naturally has a block-wise structure, we perform neural architecture search independently in each block, which largely reduces the search space. Different from previous block-wise NAS methods, DiffNAS contains a block-wise local search strategy and a retraining strategy with a joint dynamic loss. Concretely, during the search process, we block-wisely select the best subnet to avoid the unfairness brought by the global search strategy used in previous works. When retraining the searched architecture, we adopt a dynamic joint loss to maintain the consistency between supernet training and subnet retraining, which also provides informative objectives for each block and shortens the paths of gradient propagation. We demonstrate this joint loss can effectively improve model performance. We also prove the necessity of the dynamic adjustment of this loss. The experiments show that our method can achieve significant computational reduction, especially on latent diffusion models with about 50\% MACs and Parameter reduction.

Wenxun Dai, Zhiyuan Zhao, Yule Zhong et al.

Unified multimodal models (UMMs) have achieved remarkable progress yet remain constrained by a single-turn interaction paradigm, effectively functioning as solvers for independent requests rather than assistants in continuous dialogue. To bridge this gap, we present ChatUMM. As a conversational unified model, it excels at robust context tracking to sustain interleaved multimodal generation. ChatUMM derives its capabilities from two key innovations: an interleaved multi-turn training strategy that models serialized text-image streams as a continuous conversational flow, and a systematic conversational data synthesis pipeline. This pipeline transforms a diverse set of standard single-turn datasets into fluid dialogues through three progressive stages: constructing basic stateful dialogues, enforcing long-range dependency resolution via ``distractor'' turns with history-dependent query rewriting, and synthesizing naturally interleaved multimodal responses. Extensive evaluations demonstrate that ChatUMM achieves state-of-the-art performance among open-source unified models on visual understanding and instruction-guided editing benchmarks, while maintaining competitive fidelity in text-to-image generation. Notably, ChatUMM exhibits superior robustness in complex multi-turn scenarios, ensuring fluid, context-aware dialogues.

Xiao Liu, Guangyi Chen, Yansong Tang et al.

Composing simple elements into complex concepts is crucial yet challenging, especially for 3D action generation. Existing methods largely rely on extensive neural language annotations to discern composable latent semantics, a process that is often costly and labor-intensive. In this study, we introduce a novel framework to generate compositional actions without reliance on language auxiliaries. Our approach consists of three main components: Action Coupling, Conditional Action Generation, and Decoupling Refinement. Action Coupling utilizes an energy model to extract the attention masks of each sub-action, subsequently integrating two actions using these attentions to generate pseudo-training examples. Then, we employ a conditional generative model, CVAE, to learn a latent space, facilitating the diverse generation. Finally, we propose Decoupling Refinement, which leverages a self-supervised pre-trained model MAE to ensure semantic consistency between the sub-actions and compositional actions. This refinement process involves rendering generated 3D actions into 2D space, decoupling these images into two sub-segments, using the MAE model to restore the complete image from sub-segments, and constraining the recovered images to match images rendered from raw sub-actions. Due to the lack of existing datasets containing both sub-actions and compositional actions, we created two new datasets, named HumanAct-C and UESTC-C, and present a corresponding evaluation metric. Both qualitative and quantitative assessments are conducted to show our efficacy.

Chubin Zhang, Jianan Wang, Zifeng Gao et al.

Generalist Vision-Language-Action models are currently hindered by the scarcity of robotic data compared to the abundance of human video demonstrations. Existing Latent Action Models attempt to leverage video data but often suffer from visual entanglement, capturing noise rather than manipulation skills. To address this, we propose Contrastive Latent Action Pretraining (CLAP), a framework that aligns the visual latent space from videos with a proprioceptive latent space from robot trajectories. By employing contrastive learning, CLAP maps video transitions onto a quantized, physically executable codebook. Building on this representation, we introduce a dual-formulation VLA framework offering both CLAP-NTP, an autoregressive model excelling at instruction following and object generalization, and CLAP-RF, a Rectified Flow-based policy designed for high-frequency, precise manipulation. Furthermore, we propose a Knowledge Matching (KM) regularization strategy to mitigate catastrophic forgetting during fine-tuning. Extensive experiments demonstrate that CLAP significantly outperforms strong baselines, enabling the effective transfer of skills from human videos to robotic execution. Project page: https://lin-shan.com/CLAP/.

Ruihao Xu, Yong Liu, Yansong Tang et al.

With the widespread application of drones in recent years, object detection of aerial images has attracted increasing attention, especially open-vocabulary aerial detection which is not restricted to predefined categories. Due to the scarcity of drone's viewpoint images and their significant differences from natural images, it is difficult to achieve satisfying results by directly applying vanilla open-vocabulary detection methods designed for natural scenarios. Some studies propose to transfer knowledge from pre-trained models by using lightweight networks or generating pseudo labels, but they tend to rely on models trained on natural images, neglecting the potential of foundation models specifically tailored for remote sensing and aerial imagery. To address this limitation, we propose DisDop, a unified framework that systematically distills multi-level domain priors from remote sensing foundation models (e.g., RemoteCLIP and DINOv3) into a lightweight detector. Specifically, we first distill visual priors through a teacher fusion strategy that combines RemoteCLIP's cross-modal alignment capability with DINOv3's fine-grained local feature extraction ability, transferring their complementary strengths to the detector's backbone. Second, we distill textual priors embedded in RemoteCLIP's text encoder by explicitly modeling inter-category semantic relationships, while incorporating global contextual priors to enhance local feature representation for small objects. Through this multi-level prior distillation framework, our DisDop achieves new state-of-the-art performance on open-vocabulary aerial detection benchmarks. Extensive ablation analysis also demonstrates the rationality and effectiveness of our proposed modules.

Ruihao Xu, Xingming Shui, Jingxuan Niu et al.

As AI-powered compliance monitoring becomes increasingly important in public governance and industrial safety, the ability to provide verifiable evidence and traceable accountability signals is essential. However, existing video anomaly detection datasets focus on event-level binary classification, lacking the rule-driven, explainable analysis required for real-world compliance scenarios. We introduce FoodMonitor, a benchmark for explainable compliance analysis in commercial kitchen surveillance. FoodMonitor comprises 477 video clips with 3,307 violation annotations across a dual-channel design covering both person-level and environment-level violations. Each annotation specifies which rule was violated, what non-compliant behavior occurred, and who committed it with frame-level bounding boxes. We establish a unified evaluation protocol with a two-stage matching mechanism that separately assesses spatial localization and semantic understanding, along with a composite metric ($C_{\text{score}}$) that balances environment and person detection performance. Systematic evaluation of several state-of-the-art multimodal large language models reveals that the best-performing model achieves only 0.360 $C_{\text{score}}$, with spatial localization and fine-grained rule understanding emerging as the primary bottlenecks. Our analysis identifies two distinct failure modes: localization-dominated errors and semantics-dominated errors, providing diagnostic insights for future model development.

Ruihao Xu, Yong Liu, Yansong Tang

Food defect detection is critical for automated quality control, yet existing studies lack unified benchmarks and suffer from data scarcity. We introduce FDD-48, a comprehensive dataset with fine-grained annotations across 13 food types and 48 defect categories under diverse real-world conditions. To improve detection with limited labeled data, we propose FDDet, a semi-supervised framework featuring two key components: (1) BBoxMixUp, a data augmentation technique that mixes same-category defect regions to reduce spurious feature associations, and (2) CGPC (Consistency-Guided Pseudo-Label Calibration), which filters pseudo-labels based on intra-sample consistency. Experiments show FDDet significantly outperforms mainstream detectors on FDD-48, demonstrating its effectiveness for food defect detection under data-limited scenarios.

Shiyi Zhang, Wenxun Dai, Sujia Wang et al. · tsinghua

Action quality assessment (AQA) has become an emerging topic since it can be extensively applied in numerous scenarios. However, most existing methods and datasets focus on single-person short-sequence scenes, hindering the application of AQA in more complex situations. To address this issue, we construct a new multi-person long-form video dataset for action quality assessment named LOGO. Distinguished in scenario complexity, our dataset contains 200 videos from 26 artistic swimming events with 8 athletes in each sample along with an average duration of 204.2 seconds. As for richness in annotations, LOGO includes formation labels to depict group information of multiple athletes and detailed annotations on action procedures. Furthermore, we propose a simple yet effective method to model relations among athletes and reason about the potential temporal logic in long-form videos. Specifically, we design a group-aware attention module, which can be easily plugged into existing AQA methods, to enrich the clip-wise representations based on contextual group information. To benchmark LOGO, we systematically conduct investigations on the performance of several popular methods in AQA and action segmentation. The results reveal the challenges our dataset brings. Extensive experiments also show that our approach achieves state-of-the-art on the LOGO dataset. The dataset and code will be released at \url{https://github.com/shiyi-zh0408/LOGO }.

Kang Ge, Chen Wang, Yutao Guo et al.

Large-scale foundation models have become the mainstream deep learning method, while in civil engineering, the scale of AI models is strictly limited. In this work, a vision foundation model is introduced for crack segmentation. Two parameter-efficient fine-tuning methods, adapter and low-rank adaptation, are adopted to fine-tune the foundation model in semantic segmentation: the Segment Anything Model (SAM). The fine-tuned CrackSAM shows excellent performance on different scenes and materials. To test the zero-shot performance of the proposed method, two unique datasets related to road and exterior wall cracks are collected, annotated and open-sourced, for a total of 810 images. Comparative experiments are conducted with twelve mature semantic segmentation models. On datasets with artificial noise and previously unseen datasets, the performance of CrackSAM far exceeds that of all state-of-the-art models. CrackSAM exhibits remarkable superiority, particularly under challenging conditions such as dim lighting, shadows, road markings, construction joints, and other interference factors. These cross-scenario results demonstrate the outstanding zero-shot capability of foundation models and provide new ideas for developing vision models in civil engineering.

Haonan Jiang, Yuji Wang, Yongjie Zhu et al.

Leveraging Multimodal Large Language Models (MLLMs) has become pivotal for advancing Universal Multimodal Embeddings (UME) in addressing diverse cross-modal tasks. Recent studies demonstrate that incorporating generative Chain-of-Thought (CoT) reasoning can substantially enhance task-specific representations compared to discriminative methods. However, the generated reasoning CoTs of existing generative embedding methods are limited to the textual analysis of queries and are irrelevant to the retrieval of the targets. To address these limitations, we propose a reasoning-driven UME framework that integrates Embedder-Guided Reinforcement Learning (EG-RL) to optimize the Reasoner to produce evidential Traceability CoT (T-CoT). Our key contributions are threefold: (1) We design an EG-RL framework where the Embedder provides explicit supervision to the Reasoner, ensuring the generated CoT traces are aligned with embedding tasks. (2) We introduce T-CoT, which extracts critical multimodal cues to focus on retrieval-relevant elements and provides multimodal inputs for the Embedder. (3) With limited computational resources, our framework outperforms the pioneering embedding model on both MMEB-V2 and UVRB benchmarks. The integration of multimodal evidence in structured reasoning, paired with retrieval-oriented alignment, effectively strengthens cross-modal semantic consistency and boosts the fine-grained matching capability of the model as well as the generalization across complex scenarios. Our work demonstrates that targeted reasoning optimization can significantly improve multimodal embedding quality, providing a practical and efficient solution for reasoning-driven UME development.

Shiyi Zhang, Sule Bai, Guangyi Chen et al.

In this paper, we investigate a new problem called narrative action evaluation (NAE). NAE aims to generate professional commentary that evaluates the execution of an action. Unlike traditional tasks such as score-based action quality assessment and video captioning involving superficial sentences, NAE focuses on creating detailed narratives in natural language. These narratives provide intricate descriptions of actions along with objective evaluations. NAE is a more challenging task because it requires both narrative flexibility and evaluation rigor. One existing possible solution is to use multi-task learning, where narrative language and evaluative information are predicted separately. However, this approach results in reduced performance for individual tasks because of variations between tasks and differences in modality between language information and evaluation information. To address this, we propose a prompt-guided multimodal interaction framework. This framework utilizes a pair of transformers to facilitate the interaction between different modalities of information. It also uses prompts to transform the score regression task into a video-text matching task, thus enabling task interactivity. To support further research in this field, we re-annotate the MTL-AQA and FineGym datasets with high-quality and comprehensive action narration. Additionally, we establish benchmarks for NAE. Extensive experiment results prove that our method outperforms separate learning methods and naive multi-task learning methods. Data and code are released at https://github.com/shiyi-zh0408/NAE_CVPR2024.

Ao Li, Yonggen Ling, Yiyang Lin et al.

Accurate 3D human keypoints localization is a critical technology enabling robots to achieve natural and safe physical interaction with users. Conventional 3D human keypoints estimation methods primarily focus on the whole-body reconstruction quality relative to the root joint. However, in practical human-robot interaction (HRI) scenarios, robots are more concerned with the precise metric-scale spatial localization of task-relevant body parts under the egocentric camera 3D coordinate. We propose TAIHRI, the first Vision-Language Model (VLM) tailored for close-range HRI perception, capable of understanding users' motion commands and directing the robot's attention to the most task-relevant keypoints. By quantizing 3D keypoints into a finite interaction space, TAIHRI precisely localize the 3D spatial coordinates of critical body parts by 2D keypoint reasoning via next token prediction, and seamlessly adapt to downstream tasks such as natural language control or global space human mesh recovery. Experiments on egocentric interaction benchmarks demonstrate that TAIHRI achieves superior estimation accuracy for task-critical body parts. We believe TAIHRI opens new research avenues in the field of embodied human-robot interaction. Code is available at: https://github.com/Tencent/TAIHRI.

Sule Bai, Yong Liu, Yifei Han et al.

Recent advancements in pre-trained vision-language models like CLIP, have enabled the task of open-vocabulary segmentation. CLIP demonstrates impressive zero-shot capabilities in various downstream tasks that require holistic image understanding. However, due to its image-level pre-training, CLIP struggles to capture local details, resulting in poor performance in segmentation tasks. Our analysis reveals that anomaly tokens emerge during the forward pass, drawing excessive attention from normal patch tokens, thereby diminishing spatial awareness. To address this issue, we propose Self-Calibrated CLIP (SC-CLIP), a training-free method that calibrates CLIP to produce finer representations while preserving its original generalization ability, without introducing new parameters or relying on additional backbones. Specifically, we first identify and resolve the anomaly tokens to mitigate their negative impact. Next, we enhance feature discriminability and attention correlation by leveraging the semantic consistency found in CLIP's intermediate features. Furthermore, we explore how to effectively employ multi-level feature fusion under the training-free setting. Collectively, these strategies enhance CLIP's feature representation with greater granularity and coherence. Experimental results demonstrate the effectiveness of SC-CLIP, achieving state-of-the-art results across all datasets and surpassing previous methods by 9.5%. Notably, SC-CLIP boosts the performance of vanilla CLIP ViT-L/14 by 6.8 times. Our source code is available at https://github.com/SuleBai/SC-CLIP.

Haoji Zhang, Xin Gu, Jiawen Li et al.

The video reasoning ability of multimodal large language models (MLLMs) is crucial for downstream tasks like video question answering and temporal grounding. While recent approaches have explored text-based chain-of-thought (CoT) reasoning for MLLMs, these methods often suffer from limited cross-modal interaction and increased hallucination, especially with longer videos or reasoning chains. To address these challenges, we propose Video Intelligence via Tool-Augmented Learning (VITAL), a novel end-to-end agentic video reasoning framework. With a visual toolbox, the model can densely sample new video frames on demand and generate multimodal CoT for precise long video reasoning. We observe that temporal grounding and question answering are mutually beneficial for video understanding tasks. Therefore, we construct two high-quality multi-task video reasoning datasets MTVR-CoT-72k for supervised fine-tuning and MTVR-RL-110k for reinforcement learning. Moreover, we propose a Difficulty-aware Group Relative Policy Optimization algorithm (DGRPO) to mitigate difficulty imbalance in multi-task reinforcement learning. Extensive experiments on 11 challenging video understanding benchmarks demonstrate the advanced reasoning ability of VITAL, outperforming existing methods in video question answering and temporal grounding tasks, especially in long video scenarios. Code is available at https://zhang9302002.github.io/thinkingwithvideos-page/.

Haipeng Luo, Huawen Feng, Qingfeng Sun et al.

Large Reasoning Models (LRMs) like o3 and DeepSeek-R1 have achieved remarkable progress in natural language reasoning with long chain-of-thought. However, they remain computationally inefficient and struggle with accuracy when solving problems requiring complex mathematical operations. In this work, we present AgentMath, an agent framework that seamlessly integrates language models' reasoning capabilities with code interpreters' computational precision to efficiently tackle complex mathematical problems. Our approach introduces three key innovations: (1) An automated method that converts natural language chain-of-thought into structured tool-augmented trajectories, generating high-quality supervised fine-tuning (SFT) data to alleviate data scarcity; (2) A novel agentic reinforcement learning (RL) paradigm that dynamically interleaves natural language generation with real-time code execution. This enables models to autonomously learn optimal tool-use strategies through multi-round interactive feedback, while fostering emergent capabilities in code refinement and error correction; (3) An efficient training system incorporating innovative techniques, including request-level asynchronous rollout scheduling, agentic partial rollout, and prefix-aware weighted load balancing, achieving 4-5x speedup and making efficient RL training feasible on ultra-long sequences with scenarios with massive tool calls.Extensive evaluations show that AgentMath achieves state-of-the-art performance on challenging mathematical competition benchmarks including AIME24, AIME25, and HMMT25. Specifically, AgentMath-30B-A3B attains 90.6%, 86.4%, and 73.8% accuracy respectively, achieving advanced capabilities.These results validate the effectiveness of our approach and pave the way for building more sophisticated and scalable mathematical reasoning agents.

Yiyang Fu, Chubin Zhang, Shukai Gong et al.

It is infeasible to encompass all possible disturbances within the training dataset. This raises a critical question regarding the robustness of Vision-Language-Action (VLA) models when encountering unseen real-world visual disturbances, particularly under imperfect visual conditions. In this work, we conduct a systematic study based on recent state-of-the-art VLA models and reveal a significant performance drop when visual disturbances absent from the training data are introduced. To mitigate this issue, we propose a lightweight adapter module grounded in information theory, termed the Information Bottleneck Adapter (IB-Adapter), which selectively filters potential noise from visual inputs. Without requiring any extra data or augmentation strategies, IB-Adapter consistently improves over the baseline by an average of 30%, while adding fewer than 10M parameters, demonstrating notable efficiency and effectiveness. Furthermore, even with a 14x smaller backbone (0.5B parameters) and no pre-training on the Open X-Embodiment dataset, our model StableVLA achieves robustness competitive with 7B-scale state-of-the-art VLAs. With negligible parameter overhead (<10M), our approach maintains accuracy on long-horizon tasks and surpasses OpenPi under both synthetic and physical visual corruptions.

Yuji Wang, Haoran Xu, Yong Liu et al.

Reference Audio-Visual Segmentation (Ref-AVS) aims to provide a pixel-wise scene understanding in Language-aided Audio-Visual Scenes (LAVS). This task requires the model to continuously segment objects referred to by text and audio from a video. Previous dual-modality methods always fail due to the lack of a third modality and the existing triple-modality method struggles with spatio-temporal consistency, leading to the target shift of different frames. In this work, we introduce a novel framework, termed SAM2-LOVE, which integrates textual, audio, and visual representations into a learnable token to prompt and align SAM2 for achieving Ref-AVS in the LAVS. Technically, our approach includes a multimodal fusion module aimed at improving multimodal understanding of SAM2, as well as token propagation and accumulation strategies designed to enhance spatio-temporal consistency without forgetting historical information. We conducted extensive experiments to demonstrate that SAM2-LOVE outperforms the SOTA by 8.5\% in $\mathcal{J\&F}$ on the Ref-AVS benchmark and showcase the simplicity and effectiveness of the components. Our code will be available here.

Jiajun Dong, Chengkun Wang, Wenzhao Zheng et al.

Effective image tokenization is crucial for both multi-modal understanding and generation tasks due to the necessity of the alignment with discrete text data. To this end, existing approaches utilize vector quantization (VQ) to project pixels onto a discrete codebook and reconstruct images from the discrete representation. However, compared with the continuous latent space, the limited discrete codebook space significantly restrict the representational ability of these image tokenizers. In this paper, we propose GaussianToken: An Effective Image Tokenizer with 2D Gaussian Splatting as a solution. We first represent the encoded samples as multiple flexible featured 2D Gaussians characterized by positions, rotation angles, scaling factors, and feature coefficients. We adopt the standard quantization for the Gaussian features and then concatenate the quantization results with the other intrinsic Gaussian parameters before the corresponding splatting operation and the subsequent decoding module. In general, GaussianToken integrates the local influence of 2D Gaussian distribution into the discrete space and thus enhances the representation capability of the image tokenizer. Competitive reconstruction performances on CIFAR, Mini-ImageNet, and ImageNet-1K demonstrate the effectiveness of our framework. Our code is available at: https://github.com/ChrisDong-THU/GaussianToken.

Haoji Zhang, Yiqin Wang, Yansong Tang et al.

Benefiting from the advances in large language models and cross-modal alignment, existing multimodal large language models have achieved prominent performance in image and short video understanding. However, the understanding of long videos is still challenging, as their long-context nature results in significant computational and memory overhead. Most existing work treats long videos in the same way as short videos, which is inefficient for real-world applications and hard to generalize to even longer videos. To address these issues, we propose Flash-VStream, an efficient video language model capable of processing extremely long videos and responding to user queries in real time. Particularly, we design a Flash Memory module, containing a low-capacity context memory to aggregate long-context temporal information and model the distribution of information density, and a high-capacity augmentation memory to retrieve detailed spatial information based on this distribution. Compared to existing models, Flash-VStream achieves significant reductions in inference latency. Extensive experiments on long video benchmarks and comprehensive video benchmarks, i.e., EgoSchema, MLVU, LVBench, MVBench and Video-MME, demonstrate the state-of-the-art performance and outstanding efficiency of our method. Code is available at https://github.com/IVGSZ/Flash-VStream.

Shiyi Zhang, Junhao Zhuang, Zhaoyang Zhang et al.

Action customization involves generating videos where the subject performs actions dictated by input control signals. Current methods use pose-guided or global motion customization but are limited by strict constraints on spatial structure, such as layout, skeleton, and viewpoint consistency, reducing adaptability across diverse subjects and scenarios. To overcome these limitations, we propose FlexiAct, which transfers actions from a reference video to an arbitrary target image. Unlike existing methods, FlexiAct allows for variations in layout, viewpoint, and skeletal structure between the subject of the reference video and the target image, while maintaining identity consistency. Achieving this requires precise action control, spatial structure adaptation, and consistency preservation. To this end, we introduce RefAdapter, a lightweight image-conditioned adapter that excels in spatial adaptation and consistency preservation, surpassing existing methods in balancing appearance consistency and structural flexibility. Additionally, based on our observations, the denoising process exhibits varying levels of attention to motion (low frequency) and appearance details (high frequency) at different timesteps. So we propose FAE (Frequency-aware Action Extraction), which, unlike existing methods that rely on separate spatial-temporal architectures, directly achieves action extraction during the denoising process. Experiments demonstrate that our method effectively transfers actions to subjects with diverse layouts, skeletons, and viewpoints. We release our code and model weights to support further research at https://shiyi-zh0408.github.io/projectpages/FlexiAct/

Tengbo Yu, Guanxing Lu, Zaijia Yang et al.

Multi-task robotic bimanual manipulation is becoming increasingly popular as it enables sophisticated tasks that require diverse dual-arm collaboration patterns. Compared to unimanual manipulation, bimanual tasks pose challenges to understanding the multi-body spatiotemporal dynamics. An existing method ManiGaussian pioneers encoding the spatiotemporal dynamics into the visual representation via Gaussian world model for single-arm settings, which ignores the interaction of multiple embodiments for dual-arm systems with significant performance drop. In this paper, we propose ManiGaussian++, an extension of ManiGaussian framework that improves multi-task bimanual manipulation by digesting multi-body scene dynamics through a hierarchical Gaussian world model. To be specific, we first generate task-oriented Gaussian Splatting from intermediate visual features, which aims to differentiate acting and stabilizing arms for multi-body spatiotemporal dynamics modeling. We then build a hierarchical Gaussian world model with the leader-follower architecture, where the multi-body spatiotemporal dynamics is mined for intermediate visual representation via future scene prediction. The leader predicts Gaussian Splatting deformation caused by motions of the stabilizing arm, through which the follower generates the physical consequences resulted from the movement of the acting arm. As a result, our method significantly outperforms the current state-of-the-art bimanual manipulation techniques by an improvement of 20.2% in 10 simulated tasks, and achieves 60% success rate on average in 9 challenging real-world tasks. Our code is available at https://github.com/April-Yz/ManiGaussian_Bimanual.

Zhiheng Li, Muheng Li, Jixuan Fan et al.

Scale arbitrary super-resolution based on implicit image function gains increasing popularity since it can better represent the visual world in a continuous manner. However, existing scale arbitrary works are trained and evaluated on simulated datasets, where low-resolution images are generated from their ground truths by the simplest bicubic downsampling. These models exhibit limited generalization to real-world scenarios due to the greater complexity of real-world degradations. To address this issue, we build a RealArbiSR dataset, a new real-world super-resolution benchmark with both integer and non-integer scaling factors fo the training and evaluation of real-world scale arbitrary super-resolution. Moreover, we propose a Dual-level Deformable Implicit Representation (DDIR) to solve real-world scale arbitrary super-resolution. Specifically, we design the appearance embedding and deformation field to handle both image-level and pixel-level deformations caused by real-world degradations. The appearance embedding models the characteristics of low-resolution inputs to deal with photometric variations at different scales, and the pixel-based deformation field learns RGB differences which result from the deviations between the real-world and simulated degradations at arbitrary coordinates. Extensive experiments show our trained model achieves state-of-the-art performance on the RealArbiSR and RealSR benchmarks for real-world scale arbitrary super-resolution. The dataset and code are available at \url{https://github.com/nonozhizhiovo/RealArbiSR}.