Zejia Weng

Zejia Weng, Xitong Yang, Ang Li et al. · deepmind

Contrastive Language-Image Pretraining (CLIP) has demonstrated impressive zero-shot learning abilities for image understanding, yet limited effort has been made to investigate CLIP for zero-shot video recognition. We introduce Open-VCLIP, a simple yet effective approach that transforms CLIP into a strong zero-shot video classifier that can recognize unseen actions and events at test time. Our framework extends CLIP with minimal modifications to model spatial-temporal relationships in videos, making it a specialized video classifier, while striving for generalization. We formally show that training an Open-VCLIP is equivalent to continual learning with zero historical data. To address this problem, we propose Interpolated Weight Optimization, which utilizes the benefit of weight interpolation in both training and test time. We evaluate our method on three popular and challenging action recognition datasets following various zero-shot evaluation protocols and we demonstrate our approach outperforms state-of-the-art methods by clear margins. In particular, we achieve 87.9%, 58.3%, 81.1% zero-shot accuracy on UCF, HMDB and Kinetics-600 respectively, outperforming state-of-the-art methods by 8.3%, 7.8% and 12.2%. Code is released at https://github.com/wengzejia1/Open-VCLIP.

Junke Wang, Lingchen Meng, Zejia Weng et al.

Existing visual instruction tuning methods typically prompt large language models with textual descriptions to generate instruction-following data. Despite the promising performance achieved, these descriptions are derived from image annotations, which are oftentimes coarse-grained. Furthermore, the instructions might even contradict the visual content without observing the entire visual context. To address this challenge, we introduce a fine-grained visual instruction dataset, LVIS-Instruct4V, which contains 220K visually aligned and context-aware instructions produced by prompting the powerful GPT-4V with images from LVIS. Through experimental validation and case studies, we demonstrate that high-quality visual instructional data could improve the performance of LLaVA-1.5, a state-of-the-art large multimodal model, across a wide spectrum of benchmarks by clear margins. Notably, by simply replacing the LLaVA-Instruct with our LVIS-Instruct4V, we achieve better results than LLaVA on most challenging LMM benchmarks, e.g., LLaVA$^w$ (76.7 vs. 70.7) and MM-Vet (40.2 vs. 35.4). We release our data and model at https://github.com/X2FD/LVIS-INSTRUCT4V.

Zuxuan Wu, Zejia Weng, Wujian Peng et al.

Despite significant results achieved by Contrastive Language-Image Pretraining (CLIP) in zero-shot image recognition, limited effort has been made exploring its potential for zero-shot video recognition. This paper presents Open-VCLIP++, a simple yet effective framework that adapts CLIP to a strong zero-shot video classifier, capable of identifying novel actions and events during testing. Open-VCLIP++ minimally modifies CLIP to capture spatial-temporal relationships in videos, thereby creating a specialized video classifier while striving for generalization. We formally demonstrate that training Open-VCLIP++ is tantamount to continual learning with zero historical data. To address this problem, we introduce Interpolated Weight Optimization, a technique that leverages the advantages of weight interpolation during both training and testing. Furthermore, we build upon large language models to produce fine-grained video descriptions. These detailed descriptions are further aligned with video features, facilitating a better transfer of CLIP to the video domain. Our approach is evaluated on three widely used action recognition datasets, following a variety of zero-shot evaluation protocols. The results demonstrate that our method surpasses existing state-of-the-art techniques by significant margins. Specifically, we achieve zero-shot accuracy scores of 88.1%, 58.7%, and 81.2% on UCF, HMDB, and Kinetics-600 datasets respectively, outpacing the best-performing alternative methods by 8.5%, 8.2%, and 12.3%. We also evaluate our approach on the MSR-VTT video-text retrieval dataset, where it delivers competitive video-to-text and text-to-video retrieval performance, while utilizing substantially less fine-tuning data compared to other methods. Code is released at https://github.com/wengzejia1/Open-VCLIP.

Zejia Weng, Xitong Yang, Zhen Xing et al.

Video diffusion models are able to generate high-quality videos by learning strong spatial-temporal priors on large-scale datasets. In this paper, we aim to investigate whether such priors derived from a generative process are suitable for video recognition, and eventually joint optimization of generation and recognition. Building upon Stable Video Diffusion, we introduce GenRec, the first unified framework trained with a random-frame conditioning process so as to learn generalized spatial-temporal representations. The resulting framework can naturally supports generation and recognition, and more importantly is robust even when visual inputs contain limited information. Extensive experiments demonstrate the efficacy of GenRec for both recognition and generation. In particular, GenRec achieves competitive recognition performance, offering 75.8% and 87.2% accuracy on SSV2 and K400, respectively. GenRec also performs the best on class-conditioned image-to-video generation, achieving 46.5 and 49.3 FVD scores on SSV2 and EK-100 datasets. Furthermore, GenRec demonstrates extraordinary robustness in scenarios that only limited frames can be observed. Code will be available at https://github.com/wengzejia1/GenRec.

Ailin Huang, Chengyuan Yao, Chunrui Han et al.

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

Ailin Huang, Ang Li, Aobo Kong et al.

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

Yana Wei, Liang Zhao, Jianjian Sun et al. · tsinghua

The remarkable reasoning capability of large language models (LLMs) stems from cognitive behaviors that emerge through reinforcement with verifiable rewards. This work investigates how to transfer this principle to Multimodal LLMs (MLLMs) to unlock advanced visual reasoning. We introduce a two-stage paradigm built on Qwen2.5-VL-7B: a massive linguistic cold-start fine-tuning, followed by multimodal reinforcement learning (RL) spanning nearly 1,000 steps, surpassing all previous open-source efforts in scale. This pioneering work reveals three fundamental insights: 1) Behavior transfer emerges surprisingly early in cold start due to linguistic mental imagery. 2) Cold start broadly memorizes visual behaviors, while RL critically discerns and scales up effective patterns. 3) Transfer strategically favors high-utility behaviors such as visual reflection. Our resulting model, Open-Vision-Reasoner (OVR), achieves state-of-the-art performance on a suite of reasoning benchmarks, including 95.3% on MATH500, 51.8% on MathVision and 54.6% on MathVerse. We release our model, data, and training dynamics to catalyze the development of more capable, behavior-aligned multimodal reasoners.

Zejia Weng, Xitong Yang, Ang Li et al.

We study the training of Vision Transformers for semi-supervised image classification. Transformers have recently demonstrated impressive performance on a multitude of supervised learning tasks. Surprisingly, we show Vision Transformers perform significantly worse than Convolutional Neural Networks when only a small set of labeled data is available. Inspired by this observation, we introduce a joint semi-supervised learning framework, Semiformer, which contains a transformer stream, a convolutional stream and a carefully designed fusion module for knowledge sharing between these streams. The convolutional stream is trained on limited labeled data and further used to generate pseudo labels to supervise the training of the transformer stream on unlabeled data. Extensive experiments on ImageNet demonstrate that Semiformer achieves 75.5% top-1 accuracy, outperforming the state-of-the-art by a clear margin. In addition, we show, among other things, Semiformer is a general framework that is compatible with most modern transformer and convolutional neural architectures. Code is available at https://github.com/wengzejia1/Semiformer.

Xing Zhang, Zuxuan Wu, Zejia Weng et al.

Label distributions in real-world are oftentimes long-tailed and imbalanced, resulting in biased models towards dominant labels. While long-tailed recognition has been extensively studied for image classification tasks, limited effort has been made for video domain. In this paper, we introduce VideoLT, a large-scale long-tailed video recognition dataset, as a step toward real-world video recognition. Our VideoLT contains 256,218 untrimmed videos, annotated into 1,004 classes with a long-tailed distribution. Through extensive studies, we demonstrate that state-of-the-art methods used for long-tailed image recognition do not perform well in the video domain due to the additional temporal dimension in video data. This motivates us to propose FrameStack, a simple yet effective method for long-tailed video recognition task. In particular, FrameStack performs sampling at the frame-level in order to balance class distributions, and the sampling ratio is dynamically determined using knowledge derived from the network during training. Experimental results demonstrate that FrameStack can improve classification performance without sacrificing overall accuracy. Code and dataset are available at: https://github.com/17Skye17/VideoLT.

Xingjun Ma, Linxi Jiang, Hanxun Huang et al.

Evaluating the robustness of a defense model is a challenging task in adversarial robustness research. Obfuscated gradients have previously been found to exist in many defense methods and cause a false signal of robustness. In this paper, we identify a more subtle situation called Imbalanced Gradients that can also cause overestimated adversarial robustness. The phenomenon of imbalanced gradients occurs when the gradient of one term of the margin loss dominates and pushes the attack towards to a suboptimal direction. To exploit imbalanced gradients, we formulate a Margin Decomposition (MD) attack that decomposes a margin loss into individual terms and then explores the attackability of these terms separately via a two-stage process. We also propose a multi-targeted and ensemble version of our MD attack. By investigating 24 defense models proposed since 2018, we find that 11 models are susceptible to a certain degree of imbalanced gradients and our MD attack can decrease their robustness evaluated by the best standalone baseline attack by more than 1%. We also provide an in-depth investigation on the likely causes of imbalanced gradients and effective countermeasures. Our code is available at https://github.com/HanxunH/MDAttack.

StepFun, Bin Wang, Bojun Wang et al.

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

Xin Wang, Jie Li, Zejia Weng et al.

Vision-Language-Action (VLA) models are driving rapid progress in robotics by enabling agents to interpret multimodal inputs and execute complex, long-horizon tasks. However, their safety and robustness against adversarial attacks remain largely underexplored. In this work, we identify and formalize a critical adversarial vulnerability in which adversarial images can "freeze" VLA models and cause them to ignore subsequent instructions. This threat effectively disconnects the robot's digital mind from its physical actions, potentially inducing inaction during critical interventions. To systematically study this vulnerability, we propose FreezeVLA, a novel attack framework that generates and evaluates action-freezing attacks via min-max bi-level optimization. Experiments on three state-of-the-art VLA models and four robotic benchmarks show that FreezeVLA attains an average attack success rate of 76.2%, significantly outperforming existing methods. Moreover, adversarial images generated by FreezeVLA exhibit strong transferability, with a single image reliably inducing paralysis across diverse language prompts. Our findings expose a critical safety risk in VLA models and highlight the urgent need for robust defense mechanisms.

Haolong Yan, Jia Wang, Xin Huang et al.

Recent advances in multimodal large language models unlock unprecedented opportunities for GUI automation. However, a fundamental challenge remains: how to efficiently acquire high-quality training data while maintaining annotation reliability? We introduce a self-evolving training pipeline powered by the Calibrated Step Reward System, which converts model-generated trajectories into reliable training signals through trajectory-level calibration, achieving >90% annotation accuracy with 10-100x lower cost. Leveraging this pipeline, we introduce Step-GUI, a family of models (4B/8B) that achieves state-of-the-art GUI performance (8B: 80.2% AndroidWorld, 48.5% OSWorld, 62.6% ScreenShot-Pro) while maintaining robust general capabilities. As GUI agent capabilities improve, practical deployment demands standardized interfaces across heterogeneous devices while protecting user privacy. To this end, we propose GUI-MCP, the first Model Context Protocol for GUI automation with hierarchical architecture that combines low-level atomic operations and high-level task delegation to local specialist models, enabling high-privacy execution where sensitive data stays on-device. Finally, to assess whether agents can handle authentic everyday usage, we introduce AndroidDaily, a benchmark grounded in real-world mobile usage patterns with 3146 static actions and 235 end-to-end tasks across high-frequency daily scenarios (8B: static 89.91%, end-to-end 52.50%). Our work advances the development of practical GUI agents and demonstrates strong potential for real-world deployment in everyday digital interactions.

Zhen Xing, Qi Dai, Zejia Weng et al.

Text-guided video prediction (TVP) involves predicting the motion of future frames from the initial frame according to an instruction, which has wide applications in virtual reality, robotics, and content creation. Previous TVP methods make significant breakthroughs by adapting Stable Diffusion for this task. However, they struggle with frame consistency and temporal stability primarily due to the limited scale of video datasets. We observe that pretrained Image2Video diffusion models possess good priors for video dynamics but they lack textual control. Hence, transferring Image2Video models to leverage their video dynamic priors while injecting instruction control to generate controllable videos is both a meaningful and challenging task. To achieve this, we introduce the Multi-Modal Large Language Model (MLLM) to predict future video states based on initial frames and text instructions. More specifically, we design a dual query transformer (DQFormer) architecture, which integrates the instructions and frames into the conditional embeddings for future frame prediction. Additionally, we develop Long-Short Term Temporal Adapters and Spatial Adapters that can quickly transfer general video diffusion models to specific scenarios with minimal training costs. Experimental results show that our method significantly outperforms state-of-the-art techniques on four datasets: Something Something V2, Epic Kitchen-100, Bridge Data, and UCF-101. Notably, AID achieves 91.2% and 55.5% FVD improvements on Bridge and SSv2 respectively, demonstrating its effectiveness in various domains. More examples can be found at our website https://chenhsing.github.io/AID.

Wujian Peng, Zejia Weng, Hengduo Li et al.

Exploring a substantial amount of unlabeled data, semi-supervised learning (SSL) boosts the recognition performance when only a limited number of labels are provided. However, traditional methods assume that the data distribution is class-balanced, which is difficult to achieve in reality due to the long-tailed nature of real-world data. While the data imbalance problem has been extensively studied in supervised learning (SL) paradigms, directly transferring existing approaches to SSL is nontrivial, as prior knowledge about data distribution remains unknown in SSL. In light of this, we propose Balanced Memory Bank (BMB), a semi-supervised framework for long-tailed recognition. The core of BMB is an online-updated memory bank that caches historical features with their corresponding pseudo labels, and the memory is also carefully maintained to ensure the data therein are class-rebalanced. Additionally, an adaptive weighting module is introduced to work jointly with the memory bank so as to further re-calibrate the biased training process. We conduct experiments on multiple datasets and demonstrate, among other things, that BMB surpasses state-of-the-art approaches by clear margins, for example 8.2$\%$ on the 1$\%$ labeled subset of ImageNet127 (with a resolution of 64$\times$64) and 4.3$\%$ on the 50$\%$ labeled subset of ImageNet-LT.

Zejia Weng, Lingchen Meng, Rui Wang et al.

There is a growing trend in placing video advertisements on social platforms for online marketing, which demands automatic approaches to understand the contents of advertisements effectively. Taking the 2021 TAAC competition as an opportunity, we developed a multimodal system to improve the ability of structured analysis of advertising video content. In our framework, we break down the video structuring analysis problem into two tasks, i.e., scene segmentation and multi-modal tagging. In scene segmentation, we build upon a temporal convolution module for temporal modeling to predict whether adjacent frames belong to the same scene. In multi-modal tagging, we first compute clip-level visual features by aggregating frame-level features with NeXt-SoftDBoF. The visual features are further complemented with textual features that are derived using a global-local attention mechanism to extract useful information from OCR (Optical Character Recognition) and ASR (Audio Speech Recognition) outputs. Our solution achieved a score of 0.2470 measured in consideration of localization and prediction accuracy, ranking fourth in the 2021 TAAC final leaderboard.

Rui Wang, Zuxuan Wu, Zejia Weng et al.

Unsupervised domain adaptation (UDA) aims to transfer knowledge learned from a fully-labeled source domain to a different unlabeled target domain. Most existing UDA methods learn domain-invariant feature representations by minimizing feature distances across domains. In this work, we build upon contrastive self-supervised learning to align features so as to reduce the domain discrepancy between training and testing sets. Exploring the same set of categories shared by both domains, we introduce a simple yet effective framework CDCL, for domain alignment. In particular, given an anchor image from one domain, we minimize its distances to cross-domain samples from the same class relative to those from different categories. Since target labels are unavailable, we use a clustering-based approach with carefully initialized centers to produce pseudo labels. In addition, we demonstrate that CDCL is a general framework and can be adapted to the data-free setting, where the source data are unavailable during training, with minimal modification. We conduct experiments on two widely used domain adaptation benchmarks, i.e., Office-31 and VisDA-2017, for image classification tasks, and demonstrate that CDCL achieves state-of-the-art performance on both datasets.

Zejia Weng, Zuxuan Wu, Hengduo Li et al.

Videos are multimodal in nature. Conventional video recognition pipelines typically fuse multimodal features for improved performance. However, this is not only computationally expensive but also neglects the fact that different videos rely on different modalities for predictions. This paper introduces Hierarchical and Conditional Modality Selection (HCMS), a simple yet efficient multimodal learning framework for efficient video recognition. HCMS operates on a low-cost modality, i.e., audio clues, by default, and dynamically decides on-the-fly whether to use computationally-expensive modalities, including appearance and motion clues, on a per-input basis. This is achieved by the collaboration of three LSTMs that are organized in a hierarchical manner. In particular, LSTMs that operate on high-cost modalities contain a gating module, which takes as inputs lower-level features and historical information to adaptively determine whether to activate its corresponding modality; otherwise it simply reuses historical information. We conduct extensive experiments on two large-scale video benchmarks, FCVID and ActivityNet, and the results demonstrate the proposed approach can effectively explore multimodal information for improved classification performance while requiring much less computation.