Real-world video editing demands not only expert knowledge of cinematic techniques but also multimodal reasoning to select, align, and combine footage into coherent narratives. While recent Large Multimodal Models (LMMs) have shown remarkable progress in general video understanding, their abilities in multi-video reasoning and operational editing workflows remain largely unexplored. We introduce VEBENCH, the first comprehensive benchmark designed to evaluate both editing knowledge understanding and operational reasoning in realistic video editing scenarios. VEBENCH contains 3.9K high-quality edited videos (over 257 hours) and 3,080 human-verified QA pairs, built through a three-round human-AI collaborative annotation pipeline that ensures precise temporal labeling and semantic consistency. It features two complementary QA tasks: 1) Video Editing Technique Recognition, assessing models' ability to identify 7 editing techniques using multimodal cues; and 2) Video Editing Operation Simulation, modeling real-world editing workflows by requiring the selection and temporal localization of relevant clips from multiple candidates. Extensive experiments across proprietary (e.g., Gemini-2.5-Pro) and open-source LMMs reveal a large gap between current model performance and human-level editing cognition. These results highlight the urgent need for bridging video understanding with creative operational reasoning. We envision VEBENCH as a foundation for advancing intelligent video editing systems and driving future research on complex reasoning.

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI – effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 License at https://github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3. The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3.

Humans naturally share information with those they are connected to, and video has become one of the dominant mediums for communication and expression on the Internet. To support the creation of high-quality large-scale video content, a modern pipeline requires a comprehensive understanding of both the raw input materials (e.g., the unedited footage captured by cameras) and the editing components (e.g., visual effects). In video editing scenarios, models must process multiple modalities (e.g., vision, audio, text) with strong background knowledge and handle flexible input lengths (e.g., hour-long raw videos), which poses significant challenges for traditional models. In this report, we introduce Vidi, a family of Large Multimodal Models (LMMs) for a wide range of video understand editing scenarios. The first release focuses on temporal retrieval, i.e., identifying the time ranges within the input videos corresponding to a given text query, which plays a critical role in intelligent editing. The model is capable of processing hour-long videos with strong temporal understanding capability, e.g., retrieve time ranges for certain queries. To support a comprehensive evaluation in real-world scenarios, we also present the VUE-TR benchmark, which introduces five key advancements. 1) Video duration: significantly longer than videos of existing temporal retrival datasets, 2) Audio support: includes audio-based queries, 3) Query format: diverse query lengths/formats, 4) Annotation quality: ground-truth time ranges are manually annotated. 5) Evaluation metric: a refined IoU metric to support evaluation over multiple time ranges. Remarkably, Vidi significantly outperforms leading proprietary models, e.g., GPT-4o and Gemini, on the temporal retrieval task, indicating its superiority in video editing scenarios.

Video has emerged as the primary medium for communication and creativity on the Internet, driving strong demand for scalable, high-quality video production. Vidi models continue to evolve toward next-generation video creation and have achieved state-of-the-art performance in multimodal temporal retrieval (TR). In its second release, Vidi2 advances video understanding with fine-grained spatio-temporal grounding (STG) and extends its capability to video question answering (Video QA), enabling comprehensive multimodal reasoning. Given a text query, Vidi2 can identify not only the corresponding timestamps but also the bounding boxes of target objects within the output time ranges. To enable comprehensive evaluation of STG, we introduce a new benchmark, VUE-STG, which offers critical improvements over existing STG datasets. In addition, we upgrade the previous VUE-TR benchmark to VUE-TR-V2, achieving a more balanced duration and query distribution. Remarkably, the Vidi2 model substantially outperforms leading proprietary systems, such as Gemini 3 Pro Preview and GPT-5, on both VUE-TR-V2 and VUE-STG, while achieving competitive results with popular open-source models with similar scale on video QA benchmarks. The latest Vidi2.5 offers significantly stronger STG capability and slightly better TR and Video QA performance over Vidi2. This update also introduces a Vidi2.5-Think model to handle plot understanding with complex plot reasoning. To comprehensively evaluate the performance of plot understanding, we propose VUE-PLOT benchmark with two tracks, Character and Reasoning. Notably, Vidi2.5-Think outperforms Gemini 3 Pro Preview on fine-grained character understanding with comparable performance on complex plot reasoning. Furthermore, we demonstrate the effectiveness of Vidi2.5 on a challenging real-world application, video editing planning.

Large vision-and-language models (LVLMs) have traditionally integrated visual and textual tokens by concatenating them into a single homogeneous input for large language models (LLMs), thereby maximally preserving the pre-trained language capabilities. However, this constrained architecture for visual and textual tokens restricts the design space for processing visual tokens, potentially leading to suboptimal performance and efficiency. In this paper, we propose Decomposed Attention (D-Attn), a more flexible attention architecture for LVLMs, which enables modification of visual token operations without affecting textual-to-textual attention. D-Attn decomposes the 1-D causal self-attention of LVLMs into visual-to-visual, textual-to-visual, and textual-to-textual attentions, and the visual and textual output tokens from the decomposed attentions are merged with a carefully derived weighting strategy, namely $\alpha$-weighting. Taking advantage of the flexibility, we are able to introduce two critical improvements in visual token processing while maintaining the capacity of pre-trained LLMs: 1) We rectify the biased positional encoding in textual-to-visual attention to boost visual understanding performance. 2) We diagonalize visual-to-visual attention to reduce computation complexity from $\mathcal{O}\left(\vert V\vert^{2}\right)$ to $\mathcal{O}(\vert V\vert)$ for $\vert V\vert$ visual tokens without compromising performance. Extensive experiments and analysis validate the effectiveness of D-Attn, demonstrating significant improvements on multiple image benchmarks while significantly reducing computational costs (e.g., $5 \times$ faster). Code will be available at https://github.com/bytedance/DecomposedAttention.

Large vision-and-language models (LVLMs) have traditionally integrated visual and textual tokens by concatenating them into a single homogeneous input for large language models (LLMs), thereby maximally preserving the pre-trained language capabilities. However, this constrained architecture for visual and textual tokens restricts the design space for processing visual tokens, potentially leading to suboptimal performance and efficiency. In this paper, we propose Decomposed Attention (D-Attn), a more flexible attention architecture for LVLMs, which enables modification of visual token operations without affecting textual-to-textual attention. D-Attn decomposes the 1-D causal self-attention of LVLMs into visual-to-visual, textual-to-visual, and textual-to-textual attentions, and the visual and textual output tokens from the decomposed attentions are merged with a carefully derived weighting strategy, namely α-weighting. Taking advantage of the flexibility, we are able to introduce two critical improvements in visual token processing while maintaining the capacity of pre-trained LLMs: 1) We rectify the biased positional encoding in textual-to-visual attention to boost visual understanding performance. 2) We diagonalize visual-to-visual attention to reduce computation complexity from O(|V|^2) to O(|V|) for |V| visual tokens without compromising performance. Extensive experiments and analysis validate the effectiveness of D-Attn, demonstrating significant improvements on multiple image benchmarks while significantly reducing computational costs (, 5× faster). Code will be available at https://github.com/bytedance/DecomposedAttention.

Large Vision-Language Models (LVLMs) have shown promising performance in vision-language understanding and reasoning tasks. However, their visual understanding behaviors remain underexplored. A fundamental question arises: to what extent do LVLMs rely on visual input, and which image regions contribute to their responses? It is non-trivial to interpret the free-form generation of LVLMs due to their complicated visual architecture (e.g., multiple encoders and multi-resolution) and variable-length outputs. In this paper, we extend existing heatmap visualization methods (e.g., iGOS++) to support LVLMs for open-ended visual question answering. We propose a method to select visually relevant tokens that reflect the relevance between generated answers and input image. Furthermore, we conduct a comprehensive analysis of state-of-the-art LVLMs on benchmarks designed to require visual information to answer. Our findings offer several insights into LVLM behavior, including the relationship between focus region and answer correctness, differences in visual attention across architectures, and the impact of LLM scale on visual understanding. The code and data are available at https://github.com/bytedance/LVLM_Interpretation.

The emerging video LMMs (Large Multimodal Models) have achieved significant improvements on generic video understanding in the form of VQA (Visual Question Answering), where the raw videos are captured by cameras. However, a large portion of videos in real-world applications are edited videos, e.g., users usually cut and add effects/modifications to the raw video before publishing it on social media platforms. The edited videos usually have high view counts but they are not covered in existing benchmarks of video LMMs, i.e., ActivityNet-QA, or VideoChatGPT benchmark. In this paper, we leverage the edited videos on a popular short video platform, i.e., TikTok, and build a video VQA benchmark (named EditVid-QA) covering four typical editing categories, i.e., effect, funny, meme, and game. Funny and meme videos benchmark nuanced understanding and high-level reasoning, while effect and game evaluate the understanding capability of artificial design. Most of the open-source video LMMs perform poorly on the EditVid-QA benchmark, indicating a huge domain gap between edited short videos on social media and regular raw videos. To improve the generalization ability of LMMs, we collect a training set for the proposed benchmark based on both Panda-70M/WebVid raw videos and small-scale TikTok/CapCut edited videos, which boosts the performance on the proposed EditVid-QA benchmark, indicating the effectiveness of high-quality training data. We also identified a serious issue in the existing evaluation protocol using the GPT-3.5 judge, namely a "sorry" attack, where a sorry-style naive answer can achieve an extremely high rating from the GPT judge, e.g., over 4.3 for correctness score on VideoChatGPT evaluation protocol. To avoid the "sorry" attacks, we evaluate results with GPT-4 judge and keyword filtering. The dataset is released at https://github.com/XenonLamb/EditVid-QA.

Recent advancements in Multimodal Large Language Models (LLMs) have focused primarily on scaling by increasing text-image pair data and enhancing LLMs to improve performance on multimodal tasks. However, these scaling approaches are computationally expensive and overlook the significance of improving model capabilities from the vision side. Inspired by the successful applications of Mixture-of-Experts (MoE) in LLMs, which improves model scalability during training while keeping inference costs similar to those of smaller models, we propose CuMo. CuMo incorporates Co-upcycled Top-K sparsely-gated Mixture-of-experts blocks into both the vision encoder and the MLP connector, thereby enhancing the multimodal LLMs with minimal additional activated parameters during inference. CuMo first pre-trains the MLP blocks and then initializes each expert in the MoE block from the pre-trained MLP block during the visual instruction tuning stage. Auxiliary losses are used to ensure a balanced loading of experts. CuMo outperforms state-of-the-art multimodal LLMs across various VQA and visual-instruction-following benchmarks using models within each model size group, all while training exclusively on open-sourced datasets. The code and model weights for CuMo are open-sourced at https://github.com/SHI-Labs/CuMo.

Recent advancements in vision and language foundation models have enabled interaction with visual input through image-based queries and improved human-machine interaction and robotics applications. However, this paper argues that relying solely on visual input from pre-trained models may not suffice, with some tasks demanding additional knowledge. Therefore, it focuses on augmenting pre-trained VL foundation models with external knowledge to enhance performance and applicability. The paper introduces a novel approach that constructs a diverse knowledge database, clusters word knowledge, and integrates it into VL models, such as InstructBLIP, demonstrating improved results in open-ended multi-modal tasks. In summary, the paper contributes an innovative method for knowledge augmentation in VL models, enhancing their performance and applicability to a range of tasks.

A great deal of progress has been made in image captioning, driven by research into how to encode the image using pre-trained models. This includes visual encodings (e.g. image grid features or detected objects) and more recently textual encodings (e.g. image tags or text descriptions of image regions). As more advanced encodings are available and incorporated, it is natural to ask: how to efficiently and effectively leverage the heterogeneous set of encodings? In this paper, we propose to regard the encodings as augmented views of the input image. The image captioning model encodes each view independently with a shared encoder efficiently, and a contrastive loss is incorporated across the encoded views in a novel way to improve their representation quality and the model's data efficiency. Our proposed hierarchical decoder then adaptively weighs the encoded views according to their effectiveness for caption generation by first aggregating within each view at the token level, and then across views at the view level. We demonstrate significant performance improvements of +5.6% CIDEr on MS-COCO and +12.9% CIDEr on Flickr30k compared to state of the arts, and conduct rigorous analyses to demonstrate the importance of each part of our design.

A great deal of progress has been made in image captioning, driven by research into how to encode the image using pre-trained models. This includes visual encodings (e.g. image grid features or detected objects) and more recently textual encodings (e.g. image tags or text descriptions of image regions). As more advanced encodings are available and incorporated, it is natural to ask: how to efficiently and effectively leverage and ensemble the heterogeneous set of encodings? In this paper, we propose to regard the encodings as augmented views of the input image. The model encodes each view independently with a shared encoder efficiently, and a contrastive loss is incorporated across the encoded views to improve the representation quality, as well as to enable semi-supervised training of image captioning. Our proposed hierarchical decoder then adaptively ensembles the encoded views according to their usefulness by first ensembling within each view at the token level, and then across views at the view level. We demonstrate significant performance improvements of +5.6% CIDEr on MS-COCO compared to state of the art under the same trained-from-scratch setting and +16.8% CIDEr on Flickr30K with semi-supervised training, and conduct rigorous analyses to demonstrate the importance of each part of our design.

In Vision-and-Language Navigation (VLN), researchers typically take an image encoder pre-trained on ImageNet without fine-tuning on the environments that the agent will be trained or tested on. However, the distribution shift between the training images from ImageNet and the views in the navigation environments may render the ImageNet pre-trained image encoder suboptimal. Therefore, in this paper, we design a set of structure-encoding auxiliary tasks (SEA) that leverage the data in the navigation environments to pre-train and improve the image encoder. Specifically, we design and customize (1) 3D jigsaw, (2) traversability prediction, and (3) instance classification to pre-train the image encoder. Through rigorous ablations, our SEA pre-trained features are shown to better encode structural information of the scenes, which ImageNet pre-trained features fail to properly encode but is crucial for the target navigation task. The SEA pre-trained features can be easily plugged into existing VLN agents without any tuning. For example, on Test-Unseen environments, the VLN agents combined with our SEA pre-trained features achieve absolute success rate improvement of 12% for Speaker-Follower [14], 5% for Env-Dropout [37], and 4% for AuxRN [50].

Generalized Category Discovery (GCD) requires a model to both classify known categories and cluster unknown categories in unlabeled data. Prior methods leveraged self-supervised pre-training combined with supervised fine-tuning on the labeled data, followed by simple clustering methods. In this paper, we posit that such methods are still prone to poor performance on out-of-distribution categories, and do not leverage a key ingredient: Semantic relationships between object categories. We therefore propose to leverage multi-modal (vision and language) models, in two complementary ways. First, we establish a strong baseline by replacing uni-modal features with CLIP, inspired by its zero-shot performance. Second, we propose a novel retrieval-based mechanism that leverages CLIP's aligned vision-language representations by mining text descriptions from a text corpus for the labeled and unlabeled set. We specifically use the alignment between CLIP's visual encoding of the image and textual encoding of the corpus to retrieve top-k relevant pieces of text and incorporate their embeddings to perform joint image+text semi-supervised clustering. We perform rigorous experimentation and ablations (including on where to retrieve from, how much to retrieve, and how to combine information), and validate our results on several datasets including out-of-distribution domains, demonstrating state-of-art results.

Significant progress has been made on visual captioning, largely relying on pre-trained features and later fixed object detectors that serve as rich inputs to auto-regressive models. A key limitation of such methods, however, is that the output of the model is conditioned only on the object detector's outputs. The assumption that such outputs can represent all necessary information is unrealistic, especially when the detector is transferred across datasets. In this work, we reason about the graphical model induced by this assumption, and propose to add an auxiliary input to represent missing information such as object relationships. We specifically propose to mine attributes and relationships from the Visual Genome dataset and condition the captioning model on them. Crucially, we propose (and show to be important) the use of a multi-modal pre-trained model (CLIP) to retrieve such contextual descriptions. Further, object detector models are frozen and do not have sufficient richness to allow the captioning model to properly ground them. As a result, we propose to condition both the detector and description outputs on the image, and show qualitatively and quantitatively that this can improve grounding. We validate our method on image captioning, perform thorough analyses of each component and importance of the pre-trained multi-modal model, and demonstrate significant improvements over the current state of the art, specifically +7.5% in CIDEr and +1.3% in BLEU-4 metrics.

With the recent development of Semi-Supervised Object Detection (SS-OD) techniques, object detectors can be improved by using a limited amount of labeled data and abundant unlabeled data. However, there are still two challenges that are not addressed: (1) there is no prior SS-OD work on anchor-free detectors, and (2) prior works are ineffective when pseudo-labeling bounding box regression. In this paper, we present Unbiased Teacher v2, which shows the generalization of SS-OD method to anchor-free detectors and also introduces Listen2Student mechanism for the unsupervised regression loss. Specifically, we first present a study examining the effectiveness of existing SS-OD methods on anchor-free detectors and find that they achieve much lower performance improvements under the semi-supervised setting. We also observe that box selection with centerness and the localization-based labeling used in anchor-free detectors cannot work well under the semi-supervised setting. On the other hand, our Listen2Student mechanism explicitly prevents misleading pseudo-labels in the training of bounding box regression; we specifically develop a novel pseudo-labeling selection mechanism based on the Teacher and Student's relative uncertainties. This idea contributes to favorable improvement in the regression branch in the semi-supervised setting. Our method, which works for both anchor-free and anchor-based methods, consistently performs favorably against the state-of-the-art methods in VOC, COCO-standard, and COCO-additional.

Recent state-of-the-art semi-supervised learning (SSL) methods use a combination of image-based transformations and consistency regularization as core components. Such methods, however, are limited to simple transformations such as traditional data augmentation or convex combinations of two images. In this paper, we propose a novel learned feature-based refinement and augmentation method that produces a varied set of complex transformations. Importantly, these transformations also use information from both within-class and across-class prototypical representations that we extract through clustering. We use features already computed across iterations by storing them in a memory bank, obviating the need for significant extra computation. These transformations, combined with traditional image-based augmentation, are then used as part of the consistency-based regularization loss. We demonstrate that our method is comparable to current state of art for smaller datasets (CIFAR-10 and SVHN) while being able to scale up to larger datasets such as CIFAR-100 and mini-Imagenet where we achieve significant gains over the state of art (\textit{e.g.,} absolute 17.44\% gain on mini-ImageNet). We further test our method on DomainNet, demonstrating better robustness to out-of-domain unlabeled data, and perform rigorous ablations and analysis to validate the method.

In this paper, we propose the problem of collaborative perception, where robots can combine their local observations with those of neighboring agents in a learnable way to improve accuracy on a perception task. Unlike existing work in robotics and multi-agent reinforcement learning, we formulate the problem as one where learned information must be shared across a set of agents in a bandwidth-sensitive manner to optimize for scene understanding tasks such as semantic segmentation. Inspired by networking communication protocols, we propose a multi-stage handshake communication mechanism where the neural network can learn to compress relevant information needed for each stage. Specifically, a target agent with degraded sensor data sends a compressed request, the other agents respond with matching scores, and the target agent determines who to connect with (i.e., receive information from). We additionally develop the AirSim-CP dataset and metrics based on the AirSim simulator where a group of aerial robots perceive diverse landscapes, such as roads, grasslands, buildings, etc. We show that for the semantic segmentation task, our handshake communication method significantly improves accuracy by approximately 20% over decentralized baselines, and is comparable to centralized ones using a quarter of the bandwidth.

3D reconstruction, inferring 3D shape information from a single 2D image, has drawn attention from learning and vision communities. In this paper, we propose a framework for learning pose-aware 3D shape reconstruction. Our proposed model learns deep representation for recovering the 3D object, with the ability to extract camera pose information but without any direct supervision of ground truth camera pose. This is realized by exploitation of 2D-3D self-consistency between 2D masks and 3D voxels. Experiments qualitatively and quantitatively demonstrate the effectiveness and robustness of our model, which performs favorably against state-of-the-art methods.