Zhiqi Li

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki et al. · nvidia

We introduce Nemotron Nano V2 VL, the latest model of the Nemotron vision-language series designed for strong real-world document understanding, long video comprehension, and reasoning tasks. Nemotron Nano V2 VL delivers significant improvements over our previous model, Llama-3.1-Nemotron-Nano-VL-8B, across all vision and text domains through major enhancements in model architecture, datasets, and training recipes. Nemotron Nano V2 VL builds on Nemotron Nano V2, a hybrid Mamba-Transformer LLM, and innovative token reduction techniques to achieve higher inference throughput in long document and video scenarios. We are releasing model checkpoints in BF16, FP8, and FP4 formats and sharing large parts of our datasets, recipes and training code.

Wenhai Wang, Jifeng Dai, Zhe Chen et al.

Compared to the great progress of large-scale vision transformers (ViTs) in recent years, large-scale models based on convolutional neural networks (CNNs) are still in an early state. This work presents a new large-scale CNN-based foundation model, termed InternImage, which can obtain the gain from increasing parameters and training data like ViTs. Different from the recent CNNs that focus on large dense kernels, InternImage takes deformable convolution as the core operator, so that our model not only has the large effective receptive field required for downstream tasks such as detection and segmentation, but also has the adaptive spatial aggregation conditioned by input and task information. As a result, the proposed InternImage reduces the strict inductive bias of traditional CNNs and makes it possible to learn stronger and more robust patterns with large-scale parameters from massive data like ViTs. The effectiveness of our model is proven on challenging benchmarks including ImageNet, COCO, and ADE20K. It is worth mentioning that InternImage-H achieved a new record 65.4 mAP on COCO test-dev and 62.9 mIoU on ADE20K, outperforming current leading CNNs and ViTs. The code will be released at https://github.com/OpenGVLab/InternImage.

Aaron Blakeman, Aarti Basant, Abhinav Khattar et al. · nvidia

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3$\times$ faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. We are releasing Nemotron-H base model checkpoints with support in Hugging Face and NeMo.

Hongyang Li, Chonghao Sima, Jifeng Dai et al.

Learning powerful representations in bird's-eye-view (BEV) for perception tasks is trending and drawing extensive attention both from industry and academia. Conventional approaches for most autonomous driving algorithms perform detection, segmentation, tracking, etc., in a front or perspective view. As sensor configurations get more complex, integrating multi-source information from different sensors and representing features in a unified view come of vital importance. BEV perception inherits several advantages, as representing surrounding scenes in BEV is intuitive and fusion-friendly; and representing objects in BEV is most desirable for subsequent modules as in planning and/or control. The core problems for BEV perception lie in (a) how to reconstruct the lost 3D information via view transformation from perspective view to BEV; (b) how to acquire ground truth annotations in BEV grid; (c) how to formulate the pipeline to incorporate features from different sources and views; and (d) how to adapt and generalize algorithms as sensor configurations vary across different scenarios. In this survey, we review the most recent works on BEV perception and provide an in-depth analysis of different solutions. Moreover, several systematic designs of BEV approach from the industry are depicted as well. Furthermore, we introduce a full suite of practical guidebook to improve the performance of BEV perception tasks, including camera, LiDAR and fusion inputs. At last, we point out the future research directions in this area. We hope this report will shed some light on the community and encourage more research effort on BEV perception. We keep an active repository to collect the most recent work and provide a toolbox for bag of tricks at https://github.com/OpenDriveLab/Birds-eye-view-Perception

Johan Bjorck, Fernando Castañeda, Nikita Cherniadev et al. · nvidia

General-purpose robots need a versatile body and an intelligent mind. Recent advancements in humanoid robots have shown great promise as a hardware platform for building generalist autonomy in the human world. A robot foundation model, trained on massive and diverse data sources, is essential for enabling the robots to reason about novel situations, robustly handle real-world variability, and rapidly learn new tasks. To this end, we introduce GR00T N1, an open foundation model for humanoid robots. GR00T N1 is a Vision-Language-Action (VLA) model with a dual-system architecture. The vision-language module (System 2) interprets the environment through vision and language instructions. The subsequent diffusion transformer module (System 1) generates fluid motor actions in real time. Both modules are tightly coupled and jointly trained end-to-end. We train GR00T N1 with a heterogeneous mixture of real-robot trajectories, human videos, and synthetically generated datasets. We show that our generalist robot model GR00T N1 outperforms the state-of-the-art imitation learning baselines on standard simulation benchmarks across multiple robot embodiments. Furthermore, we deploy our model on the Fourier GR-1 humanoid robot for language-conditioned bimanual manipulation tasks, achieving strong performance with high data efficiency.

Zhiqi Li, Zhiding Yu, David Austin et al.

This technical report summarizes the winning solution for the 3D Occupancy Prediction Challenge, which is held in conjunction with the CVPR 2023 Workshop on End-to-End Autonomous Driving and CVPR 23 Workshop on Vision-Centric Autonomous Driving Workshop. Our proposed solution FB-OCC builds upon FB-BEV, a cutting-edge camera-based bird's-eye view perception design using forward-backward projection. On top of FB-BEV, we further study novel designs and optimization tailored to the 3D occupancy prediction task, including joint depth-semantic pre-training, joint voxel-BEV representation, model scaling up, and effective post-processing strategies. These designs and optimization result in a state-of-the-art mIoU score of 54.19% on the nuScenes dataset, ranking the 1st place in the challenge track. Code and models will be released at: https://github.com/NVlabs/FB-BEV.

Zhiqi Li, Zhiding Yu, Wenhai Wang et al.

View Transformation Module (VTM), where transformations happen between multi-view image features and Bird-Eye-View (BEV) representation, is a crucial step in camera-based BEV perception systems. Currently, the two most prominent VTM paradigms are forward projection and backward projection. Forward projection, represented by Lift-Splat-Shoot, leads to sparsely projected BEV features without post-processing. Backward projection, with BEVFormer being an example, tends to generate false-positive BEV features from incorrect projections due to the lack of utilization on depth. To address the above limitations, we propose a novel forward-backward view transformation module. Our approach compensates for the deficiencies in both existing methods, allowing them to enhance each other to obtain higher quality BEV representations mutually. We instantiate the proposed module with FB-BEV, which achieves a new state-of-the-art result of 62.4% NDS on the nuScenes test set. Code and models are available at https://github.com/NVlabs/FB-BEV.

Aditi, Niket Agarwal, Arslan Ali et al.

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 https://openmdw.ai/license/1-1/ License at https://github.com/nvidia/cosmos}{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 .

Cheng Wan, Hongyuan Yu, Zhiqi Li et al.

Single Image Super-Resolution (SISR) is a crucial task in low-level computer vision, aiming to reconstruct high-resolution images from low-resolution counterparts. Conventional attention mechanisms have significantly improved SISR performance but often result in complex network structures and large number of parameters, leading to slow inference speed and large model size. To address this issue, we propose the Swift Parameter-free Attention Network (SPAN), a highly efficient SISR model that balances parameter count, inference speed, and image quality. SPAN employs a novel parameter-free attention mechanism, which leverages symmetric activation functions and residual connections to enhance high-contribution information and suppress redundant information. Our theoretical analysis demonstrates the effectiveness of this design in achieving the attention mechanism's purpose. We evaluate SPAN on multiple benchmarks, showing that it outperforms existing efficient super-resolution models in terms of both image quality and inference speed, achieving a significant quality-speed trade-off. This makes SPAN highly suitable for real-world applications, particularly in resource-constrained scenarios. Notably, we won the first place both in the overall performance track and runtime track of the NTIRE 2024 efficient super-resolution challenge. Our code and models are made publicly available at https://github.com/hongyuanyu/SPAN.

Jie Zhang, Zhiqi Li, Bo Li et al.

Traditional federated optimization methods perform poorly with heterogeneous data (ie, accuracy reduction), especially for highly skewed data. In this paper, we investigate the label distribution skew in FL, where the distribution of labels varies across clients. First, we investigate the label distribution skew from a statistical view. We demonstrate both theoretically and empirically that previous methods based on softmax cross-entropy are not suitable, which can result in local models heavily overfitting to minority classes and missing classes. Additionally, we theoretically introduce a deviation bound to measure the deviation of the gradient after local update. At last, we propose FedLC (\textbf {Fed} erated learning via\textbf {L} ogits\textbf {C} alibration), which calibrates the logits before softmax cross-entropy according to the probability of occurrence of each class. FedLC applies a fine-grained calibrated cross-entropy loss to local update by adding a pairwise label margin. Extensive experiments on federated datasets and real-world datasets demonstrate that FedLC leads to a more accurate global model and much improved performance. Furthermore, integrating other FL methods into our approach can further enhance the performance of the global model.

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki et al. · amazon-science, nvidia

We introduce Nemotron 3 Nano Omni, the latest model in the Nemotron multimodal series and the first to natively support audio inputs alongside text, images, and video. Nemotron 3 Nano Omni delivers consistent accuracy improvements over its predecessor, Nemotron Nano V2 VL, across all modalities, enabled by advances in architecture, training data and recipes. In particular, Nemotron 3 delivers leading results in real-world document understanding, long audio-video comprehension, and agentic computer use. Built on the highly efficient Nemotron 3 Nano 30B-A3B backbone, Nemotron 3 Nano Omni further incorporates innovative multimodal token-reduction techniques to deliver substantially lower inference latency and higher throughput than other models of similar size. We are releasing model checkpoints in BF16, FP8, and FP4 formats, along with portions of the training data and codebase to facilitate further research and development.

MiroMind Team, Song Bai, Lidong Bing et al.

We present MiroThinker v1.0, an open-source research agent designed to advance tool-augmented reasoning and information-seeking capabilities. Unlike previous agents that only scale up model size or context length, MiroThinker explores interaction scaling at the model level, systematically training the model to handle deeper and more frequent agent-environment interactions as a third dimension of performance improvement. Unlike LLM test-time scaling, which operates in isolation and risks degradation with longer reasoning chains, interactive scaling leverages environment feedback and external information acquisition to correct errors and refine trajectories. Through reinforcement learning, the model achieves efficient interaction scaling: with a 256K context window, it can perform up to 600 tool calls per task, enabling sustained multi-turn reasoning and complex real-world research workflows. Across four representative benchmarks-GAIA, HLE, BrowseComp, and BrowseComp-ZH-the 72B variant achieves up to 81.9%, 37.7%, 47.1%, and 55.6% accuracy respectively, surpassing previous open-source agents and approaching commercial counterparts such as GPT-5-high. Our analysis reveals that MiroThinker benefits from interactive scaling consistently: research performance improves predictably as the model engages in deeper and more frequent agent-environment interactions, demonstrating that interaction depth exhibits scaling behaviors analogous to model size and context length. These findings establish interaction scaling as a third critical dimension for building next-generation open research agents, complementing model capacity and context windows.

Yizhong Zhang, Zhiqi Li, Sicheng Xu et al.

Recent research advance has significantly improved the visual realism of immersive 3D video communication. In this work we present a method to further enhance this immersive experience by adding the hand touch capability ("remote hand clapping"). In our system, each meeting participant sits in front of a large screen with haptic feedback. The local participant can reach his hand out to the screen and perform hand clapping with the remote participant as if the two participants were only separated by a virtual glass. A key challenge in emulating the remote hand touch is the realistic rendering of the participant's hand and arm as the hand touches the screen. When the hand is very close to the screen, the RGBD data required for realistic rendering is no longer available. To tackle this challenge, we present a dual representation of the user's hand. Our dual representation not only preserves the high-quality rendering usually found in recent image-based rendering systems but also allows the hand to reach the screen. This is possible because the dual representation includes both an image-based model and a 3D geometry-based model, with the latter driven by a hand skeleton tracked by a side view camera. In addition, the dual representation provides a distance-based fusion of the image-based and 3D geometry-based models as the hand moves closer to the screen. The result is that the image-based and 3D geometry-based models mutually enhance each other, leading to realistic and seamless rendering. Our experiments demonstrate that our method provides consistent hand contact experience between remote users and improves the immersive experience of 3D video communication.

Zhiqi Li, Yiming Chen, Lingzhe Zhao et al.

We introduce MVControl, a novel neural network architecture that enhances existing pre-trained multi-view 2D diffusion models by incorporating additional input conditions, e.g. edge maps. Our approach enables the generation of controllable multi-view images and view-consistent 3D content. To achieve controllable multi-view image generation, we leverage MVDream as our base model, and train a new neural network module as additional plugin for end-to-end task-specific condition learning. To precisely control the shapes and views of generated images, we innovatively propose a new conditioning mechanism that predicts an embedding encapsulating the input spatial and view conditions, which is then injected to the network globally. Once MVControl is trained, score-distillation (SDS) loss based optimization can be performed to generate 3D content, in which process we propose to use a hybrid diffusion prior. The hybrid prior relies on a pre-trained Stable-Diffusion network and our trained MVControl for additional guidance. Extensive experiments demonstrate that our method achieves robust generalization and enables the controllable generation of high-quality 3D content. Code available at https://github.com/WU-CVGL/MVControl/.

Linyan Huang, Zhiqi Li, Chonghao Sima et al.

Current research is primarily dedicated to advancing the accuracy of camera-only 3D object detectors (apprentice) through the knowledge transferred from LiDAR- or multi-modal-based counterparts (expert). However, the presence of the domain gap between LiDAR and camera features, coupled with the inherent incompatibility in temporal fusion, significantly hinders the effectiveness of distillation-based enhancements for apprentices. Motivated by the success of uni-modal distillation, an apprentice-friendly expert model would predominantly rely on camera features, while still achieving comparable performance to multi-modal models. To this end, we introduce VCD, a framework to improve the camera-only apprentice model, including an apprentice-friendly multi-modal expert and temporal-fusion-friendly distillation supervision. The multi-modal expert VCD-E adopts an identical structure as that of the camera-only apprentice in order to alleviate the feature disparity, and leverages LiDAR input as a depth prior to reconstruct the 3D scene, achieving the performance on par with other heterogeneous multi-modal experts. Additionally, a fine-grained trajectory-based distillation module is introduced with the purpose of individually rectifying the motion misalignment for each object in the scene. With those improvements, our camera-only apprentice VCD-A sets new state-of-the-art on nuScenes with a score of 63.1% NDS.

Yun Liu, Xuefeng Yan, Zhilei Chen et al.

Self-supervised learning is attracting large attention in point cloud understanding. However, exploring discriminative and transferable features still remains challenging due to their nature of irregularity and sparsity. We propose a geometrically and adaptively masked auto-encoder for self-supervised learning on point clouds, termed \textit{PointGame}. PointGame contains two core components: GATE and EAT. GATE stands for the geometrical and adaptive token embedding module; it not only absorbs the conventional wisdom of geometric descriptors that captures the surface shape effectively, but also exploits adaptive saliency to focus on the salient part of a point cloud. EAT stands for the external attention-based Transformer encoder with linear computational complexity, which increases the efficiency of the whole pipeline. Unlike cutting-edge unsupervised learning models, PointGame leverages geometric descriptors to perceive surface shapes and adaptively mines discriminative features from training data. PointGame showcases clear advantages over its competitors on various downstream tasks under both global and local fine-tuning strategies. The code and pre-trained models will be publicly available.

Maolin Wang, Ziting Mai, Xuhui Chen et al.

Traditional network analysis focuses on single-layer networks, real-world systems often form multilayer networks with multiple relationship types. However, existing methods typically fail to capture complex inter-layer dependencies by treating layers independently or aggregating them. To address this, we propose T-GINEE (Tensor-Based Generalized Multilayer-graph Estimating Equation), a statistical regularization framework combining tensor-based generalized estimating equations with task-specific loss to model cross-network correlations explicitly. Key innovations include: (1) CP tensor decomposition capturing structural dependencies via shared latent factors; (2) a generalized estimating equation framework modeling inter-layer correlations through working covariance matrices; and (3) a flexible link function accommodating characteristics like sparsity. Our theoretical analysis establishes consistency and asymptotic normality under mild conditions. Extensive experiments on synthetic and real-world datasets validate T-GINEE's effectiveness for multilayer network analysis.

Shihao Wang, Shilong Liu, Yuanguo Kuang et al.

Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.

Wenhai Wang, Jiangwei Xie, ChuanYang Hu et al.

Large language models (LLMs) have opened up new possibilities for intelligent agents, endowing them with human-like thinking and cognitive abilities. In this work, we delve into the potential of large language models (LLMs) in autonomous driving (AD). We introduce DriveMLM, an LLM-based AD framework that can perform close-loop autonomous driving in realistic simulators. To this end, (1) we bridge the gap between the language decisions and the vehicle control commands by standardizing the decision states according to the off-the-shelf motion planning module. (2) We employ a multi-modal LLM (MLLM) to model the behavior planning module of a module AD system, which uses driving rules, user commands, and inputs from various sensors (e.g., camera, lidar) as input and makes driving decisions and provide explanations; This model can plug-and-play in existing AD systems such as Apollo for close-loop driving. (3) We design an effective data engine to collect a dataset that includes decision state and corresponding explanation annotation for model training and evaluation. We conduct extensive experiments and show that our model achieves 76.1 driving score on the CARLA Town05 Long, and surpasses the Apollo baseline by 4.7 points under the same settings, demonstrating the effectiveness of our model. We hope this work can serve as a baseline for autonomous driving with LLMs. Code and models shall be released at https://github.com/OpenGVLab/DriveMLM.

Zhiqi Li, Zhiding Yu, Shiyi Lan et al.

End-to-end autonomous driving recently emerged as a promising research direction to target autonomy from a full-stack perspective. Along this line, many of the latest works follow an open-loop evaluation setting on nuScenes to study the planning behavior. In this paper, we delve deeper into the problem by conducting thorough analyses and demystifying more devils in the details. We initially observed that the nuScenes dataset, characterized by relatively simple driving scenarios, leads to an under-utilization of perception information in end-to-end models incorporating ego status, such as the ego vehicle's velocity. These models tend to rely predominantly on the ego vehicle's status for future path planning. Beyond the limitations of the dataset, we also note that current metrics do not comprehensively assess the planning quality, leading to potentially biased conclusions drawn from existing benchmarks. To address this issue, we introduce a new metric to evaluate whether the predicted trajectories adhere to the road. We further propose a simple baseline able to achieve competitive results without relying on perception annotations. Given the current limitations on the benchmark and metrics, we suggest the community reassess relevant prevailing research and be cautious whether the continued pursuit of state-of-the-art would yield convincing and universal conclusions. Code and models are available at \url{https://github.com/NVlabs/BEV-Planner}

Jinjin He, Zhiqi Li, Sinan Wang et al.

We propose Hermite-NGP, a gradient-augmented multi-resolution hash encoding designed to enable fast and accurate computation of spatial derivatives for neural PDE solvers. Unlike existing NGP-based approaches that rely on automatic differentiation or finite differences and suffer from instability or high cost, Hermite-NGP explicitly stores function values and mixed partial derivatives at hash grid vertices, allowing fully analytic evaluation of gradients, Jacobians, and Hessians via Hermite interpolation. This design preserves the efficiency and spatial adaptivity of NGP while supporting analytic differential operators up to second order. We further introduce a multi-resolution curriculum training strategy analogous to multigrid V-cycles to enable coarse-to-fine optimization. Across a range of 2D and 3D PDE benchmarks, Hermite-NGP achieves up to approximately 20 times lower error than prior neural PDE methods, and reduces wall-clock convergence time by 2 to 10 times compared to other solvers, with per-epoch training times as low as 3.5 ms for models with up to 17M parameters.

Zhiqi Li, Guo Chen, Shilong Liu et al.

Recently, promising progress has been made by open-source vision-language models (VLMs) in bringing their capabilities closer to those of proprietary frontier models. However, most open-source models only publish their final model weights, leaving the critical details of data strategies and implementation largely opaque. In this work, we address VLM post-training from a data-centric perspective, showing the key role of data strategy in developing frontier VLMs. By studying and building our post-training data strategy from scratch, we share detailed insights into the development processes, aiming to benefit the development of competitive models for the open-source community. Our introduced data strategy, together with training recipes and model design, leads to a family of performant VLMs named Eagle2. Specifically, Eagle2-9B achieves state-of-the-art results across various multimodal benchmarks, matching certain competitive models with up to 70B parameters.

Guo Chen, Zhiqi Li, Shihao Wang et al.

We introduce Eagle 2.5, a family of frontier vision-language models (VLMs) for long-context multimodal learning. Our work addresses the challenges in long video comprehension and high-resolution image understanding, introducing a generalist framework for both tasks. The proposed training framework incorporates Automatic Degrade Sampling and Image Area Preservation, two techniques that preserve contextual integrity and visual details. The framework also includes numerous efficiency optimizations in the pipeline for long-context data training. Finally, we propose Eagle-Video-110K, a novel dataset that integrates both story-level and clip-level annotations, facilitating long-video understanding. Eagle 2.5 demonstrates substantial improvements on long-context multimodal benchmarks, providing a robust solution to the limitations of existing VLMs. Notably, our best model Eagle 2.5-8B achieves 72.4% on Video-MME with 512 input frames, matching the results of top-tier commercial model such as GPT-4o and large-scale open-source models like Qwen2.5-VL-72B and InternVL2.5-78B.

Yiming Chen, Zhiqi Li, Peidong Liu

Recent breakthroughs in text-to-image generation has shown encouraging results via large generative models. Due to the scarcity of 3D assets, it is hardly to transfer the success of text-to-image generation to that of text-to-3D generation. Existing text-to-3D generation methods usually adopt the paradigm of DreamFusion, which conducts per-asset optimization by distilling a pretrained text-to-image diffusion model. The generation speed usually ranges from several minutes to tens of minutes per 3D asset, which degrades the user experience and also imposes a burden to the service providers due to the high computational budget. In this work, we present an efficient text-to-3D generation method, which requires only around 8 $ms$ to generate a 3D asset given the text prompt on a consumer graphic card. The main insight is that we exploit the images generated by a large pre-trained text-to-image diffusion model, to supervise the training of a text conditioned 3D generative adversarial network. Once the network is trained, we are able to efficiently generate a 3D asset via a single forward pass. Our method requires no 3D training data and provides an alternative approach for efficient text-to-3D generation by distilling pre-trained image diffusion models.

Dongxu Wei, Qi Xu, Zhiqi Li et al.

3D scene generation has long been dominated by 2D multi-view or video diffusion models. This is due not only to the lack of scene-level 3D latent representation, but also to the fact that most scene-level 3D visual data exists in the form of multi-view images or videos, which are naturally compatible with 2D diffusion architectures. Typically, these 2D-based approaches degrade 3D spatial extrapolation to 2D temporal extension, which introduces two fundamental issues: (i) representing 3D scenes via 2D views leads to significant representation redundancy, and (ii) latent space rooted in 2D inherently limits the spatial consistency of the generated 3D scenes. In this paper, we propose, for the first time, to perform 3D scene generation directly within an implicit 3D latent space to address these limitations. First, we repurpose frozen 2D representation encoders to construct our 3D Representation Autoencoder (3DRAE), which grounds view-coupled 2D semantic representations into a view-decoupled 3D latent representation. This enables representing 3D scenes observed from arbitrary numbers of views--at any resolution and aspect ratio--with fixed complexity and rich semantics. Then we introduce 3D Diffusion Transformer (3DDiT), which performs diffusion modeling in this 3D latent space, achieving remarkably efficient and spatially consistent 3D scene generation while supporting diverse conditioning configurations. Moreover, since our approach directly generates a 3D scene representation, it can be decoded to images and optional point maps along arbitrary camera trajectories without requiring per-trajectory diffusion sampling pass, which is common in 2D-based approaches.

Jiahan Li, Zhiqi Li, Tong Lu

This technical report describes the methods we employed for the Driving with Language track of the CVPR 2024 Autonomous Grand Challenge. We utilized a powerful open-source multimodal model, InternVL-1.5, and conducted a full-parameter fine-tuning on the competition dataset, DriveLM-nuScenes. To effectively handle the multi-view images of nuScenes and seamlessly inherit InternVL's outstanding multimodal understanding capabilities, we formatted and concatenated the multi-view images in a specific manner. This ensured that the final model could meet the specific requirements of the competition task while leveraging InternVL's powerful image understanding capabilities. Meanwhile, we designed a simple automatic annotation strategy that converts the center points of objects in DriveLM-nuScenes into corresponding bounding boxes. As a result, our single model achieved a score of 0.6002 on the final leadboard.

Zhiqi Li, Wenhuan Li, Tengfei Wang et al.

Compositionality is critical for 3D object and scene generation, but existing part-aware 3D generation methods suffer from poor scalability due to quadratic global attention costs when increasing the number of components. In this work, we present MoCA, a compositional 3D generative model with two key designs: (1) importance-based component routing that selects top-k relevant components for sparse global attention, and (2) unimportant components compression that preserve contextual priors of unselected components while reducing computational complexity of global attention. With these designs, MoCA enables efficient, fine-grained compositional 3D asset creation with scalable number of components. Extensive experiments show MoCA outperforms baselines on both compositional object and scene generation tasks. Project page: https://lizhiqi49.github.io/MoCA

Zexi Li, Jie Lin, Zhiqi Li et al.

Federated learning (FL) involves multiple heterogeneous clients collaboratively training a global model via iterative local updates and model fusion. The generalization of FL's global model has a large gap compared with centralized training, which is its bottleneck for broader applications. In this paper, we study and improve FL's generalization through a fundamental ``connectivity'' perspective, which means how the local models are connected in the parameter region and fused into a generalized global model. The term ``connectivity'' is derived from linear mode connectivity (LMC), studying the interpolated loss landscape of two different solutions (e.g., modes) of neural networks. Bridging the gap between LMC and FL, in this paper, we leverage fixed anchor models to empirically and theoretically study the transitivity property of connectivity from two models (LMC) to a group of models (model fusion in FL). Based on the findings, we propose FedGuCci(+), improving group connectivity for better generalization. It is shown that our methods can boost the generalization of FL under client heterogeneity across various tasks (4 CV datasets and 6 NLP datasets) and model architectures (e.g., ViTs and PLMs). The code is available here: \href{https://github.com/ZexiLee/fedgucci}{\faGithub~FedGuCci Codebase}.

Yuwen Xiong, Zhiqi Li, Yuntao Chen et al.

We introduce Deformable Convolution v4 (DCNv4), a highly efficient and effective operator designed for a broad spectrum of vision applications. DCNv4 addresses the limitations of its predecessor, DCNv3, with two key enhancements: 1. removing softmax normalization in spatial aggregation to enhance its dynamic property and expressive power and 2. optimizing memory access to minimize redundant operations for speedup. These improvements result in a significantly faster convergence compared to DCNv3 and a substantial increase in processing speed, with DCNv4 achieving more than three times the forward speed. DCNv4 demonstrates exceptional performance across various tasks, including image classification, instance and semantic segmentation, and notably, image generation. When integrated into generative models like U-Net in the latent diffusion model, DCNv4 outperforms its baseline, underscoring its possibility to enhance generative models. In practical applications, replacing DCNv3 with DCNv4 in the InternImage model to create FlashInternImage results in up to 80% speed increase and further performance improvement without further modifications. The advancements in speed and efficiency of DCNv4, combined with its robust performance across diverse vision tasks, show its potential as a foundational building block for future vision models.

Zhenxin Li, Kailin Li, Shihao Wang et al.

We propose Hydra-MDP, a novel paradigm employing multiple teachers in a teacher-student model. This approach uses knowledge distillation from both human and rule-based teachers to train the student model, which features a multi-head decoder to learn diverse trajectory candidates tailored to various evaluation metrics. With the knowledge of rule-based teachers, Hydra-MDP learns how the environment influences the planning in an end-to-end manner instead of resorting to non-differentiable post-processing. This method achieves the $1^{st}$ place in the Navsim challenge, demonstrating significant improvements in generalization across diverse driving environments and conditions. More details by visiting \url{https://github.com/NVlabs/Hydra-MDP}.

Guo Chen, Yifei Huang, Jilan Xu et al.

Understanding videos is one of the fundamental directions in computer vision research, with extensive efforts dedicated to exploring various architectures such as RNN, 3D CNN, and Transformers. The newly proposed architecture of state space model, e.g., Mamba, shows promising traits to extend its success in long sequence modeling to video modeling. To assess whether Mamba can be a viable alternative to Transformers in the video understanding domain, in this work, we conduct a comprehensive set of studies, probing different roles Mamba can play in modeling videos, while investigating diverse tasks where Mamba could exhibit superiority. We categorize Mamba into four roles for modeling videos, deriving a Video Mamba Suite composed of 14 models/modules, and evaluating them on 12 video understanding tasks. Our extensive experiments reveal the strong potential of Mamba on both video-only and video-language tasks while showing promising efficiency-performance trade-offs. We hope this work could provide valuable data points and insights for future research on video understanding. Code is public: https://github.com/OpenGVLab/video-mamba-suite.

Zhiqi Li, Wenhai Wang, Hongyang Li et al.

3D visual perception tasks, including 3D detection and map segmentation based on multi-camera images, are essential for autonomous driving systems. In this work, we present a new framework termed BEVFormer, which learns unified BEV representations with spatiotemporal transformers to support multiple autonomous driving perception tasks. In a nutshell, BEVFormer exploits both spatial and temporal information by interacting with spatial and temporal space through predefined grid-shaped BEV queries. To aggregate spatial information, we design spatial cross-attention that each BEV query extracts the spatial features from the regions of interest across camera views. For temporal information, we propose temporal self-attention to recurrently fuse the history BEV information. Our approach achieves the new state-of-the-art 56.9\% in terms of NDS metric on the nuScenes \texttt{test} set, which is 9.0 points higher than previous best arts and on par with the performance of LiDAR-based baselines. We further show that BEVFormer remarkably improves the accuracy of velocity estimation and recall of objects under low visibility conditions. The code is available at \url{https://github.com/zhiqi-li/BEVFormer}.

Zhiqi Li, Wenhai Wang, Enze Xie et al.

Panoptic segmentation involves a combination of joint semantic segmentation and instance segmentation, where image contents are divided into two types: things and stuff. We present Panoptic SegFormer, a general framework for panoptic segmentation with transformers. It contains three innovative components: an efficient deeply-supervised mask decoder, a query decoupling strategy, and an improved post-processing method. We also use Deformable DETR to efficiently process multi-scale features, which is a fast and efficient version of DETR. Specifically, we supervise the attention modules in the mask decoder in a layer-wise manner. This deep supervision strategy lets the attention modules quickly focus on meaningful semantic regions. It improves performance and reduces the number of required training epochs by half compared to Deformable DETR. Our query decoupling strategy decouples the responsibilities of the query set and avoids mutual interference between things and stuff. In addition, our post-processing strategy improves performance without additional costs by jointly considering classification and segmentation qualities to resolve conflicting mask overlaps. Our approach increases the accuracy 6.2\% PQ over the baseline DETR model. Panoptic SegFormer achieves state-of-the-art results on COCO test-dev with 56.2\% PQ. It also shows stronger zero-shot robustness over existing methods. The code is released at \url{https://github.com/zhiqi-li/Panoptic-SegFormer}.

Ruo-Ze Liu, Wenhai Wang, Yanjie Shen et al.

StarCraft II (SC2) is a real-time strategy game in which players produce and control multiple units to fight against opponent's units. Due to its difficulties, such as huge state space, various action space, a long time horizon, and imperfect information, SC2 has been a research hotspot in reinforcement learning. Recently, an agent called AlphaStar (AS) has been proposed, which shows good performance, obtaining a high win rate of 99.8% against human players. We implemented a mini-scaled version of it called mini-AlphaStar (mAS) based on AS's paper and pseudocode. The difference between AS and mAS is that we substituted the hyper-parameters of AS with smaller ones for mini-scale training. Codes of mAS are all open-sourced (https://github.com/liuruoze/mini-AlphaStar) for future research.

Zhiqi Li, Duowen Chen, Yuchen Sun et al.

We propose a unified, few-step generative modeling framework based on \emph{cumulative flow maps} for long-range transport in probability space, inspired by flow-map techniques for physical transport and dynamics. At its core is a cumulative-flow abstraction that connects local, instantaneous updates with finite-time transport, enabling generative models to reason about global state transitions. This perspective yields a unified few-step framework built on cumulative transport and \revise{cumulative} parameterization that applies broadly to existing diffusion- and flow-based models without being tied to a specific prediction \revise{instantiation}. Our formulation supports few-step and even one-step generation while preserving synthesis quality, requiring only minimal changes to time embeddings and training objectives, and no increase in model capacity. We demonstrate its effectiveness across diverse tasks, including image generation, geometric distribution modeling, joint prediction, and SDF generation, with reduced inference cost.

Yan Tan, Yixiao Qian, Zhiqi Li et al.

For interface tracking of an arbitrary number of materials in two dimensions, we propose a multiphase cubic MARS method that (a) represents the topology and geometry of the interface via graphs, cycles, and cubic splines, (b) applies to any number of materials with arbitrarily complex topology and geometry, (c) maintains an $(r,h)$-regularity of the interface so that the distance between any pair of adjacent markers is within a user-specified range, (d) distributes the markers adaptively along the interface so that arcs with high curvature are resolved by densely populated markers, and (e) achieves fourth-, sixth-, and eighth-order accuracy both in time and in space.} In particular, all possible types of junctions, which pose challenges to VOF methods and level-set methods, are handled with ease. Results of a variety of benchmark tests confirm the analysis and demonstrate the superior accuracy, efficiency, and versatility of the proposed method.

Zhiqi Li, Yiming Chen, Lingzhe Zhao et al.

While text-to-3D and image-to-3D generation tasks have received considerable attention, one important but under-explored field between them is controllable text-to-3D generation, which we mainly focus on in this work. To address this task, 1) we introduce Multi-view ControlNet (MVControl), a novel neural network architecture designed to enhance existing pre-trained multi-view diffusion models by integrating additional input conditions, such as edge, depth, normal, and scribble maps. Our innovation lies in the introduction of a conditioning module that controls the base diffusion model using both local and global embeddings, which are computed from the input condition images and camera poses. Once trained, MVControl is able to offer 3D diffusion guidance for optimization-based 3D generation. And, 2) we propose an efficient multi-stage 3D generation pipeline that leverages the benefits of recent large reconstruction models and score distillation algorithm. Building upon our MVControl architecture, we employ a unique hybrid diffusion guidance method to direct the optimization process. In pursuit of efficiency, we adopt 3D Gaussians as our representation instead of the commonly used implicit representations. We also pioneer the use of SuGaR, a hybrid representation that binds Gaussians to mesh triangle faces. This approach alleviates the issue of poor geometry in 3D Gaussians and enables the direct sculpting of fine-grained geometry on the mesh. Extensive experiments demonstrate that our method achieves robust generalization and enables the controllable generation of high-quality 3D content. Project page: https://lizhiqi49.github.io/MVControl/.

Zhiqi Li, Bo Zhu

This paper provides a reinterpretation of the Drifting Model~\cite{deng2026generative} through a semigroup-consistent long-short flow-map factorization. We show that a global transport process can be decomposed into a long-horizon flow map followed by a short-time terminal flow map admitting a closed-form optimal velocity representation, and that taking the terminal interval length to zero recovers exactly the drifting field together with a conservative impulse term required for flow-map consistency. Based on this perspective, we propose a new likelihood learning formulation that aligns the long-short flow-map decomposition with density evolution under transport. We validate the framework through both theoretical analysis and empirical evaluations on benchmark tests, and further provide a theoretical interpretation of the feature-space optimization while highlighting several open problems for future study.

Dongxu Wei, Zhiqi Li, Peidong Liu

Prior works employing pixel-based Gaussian representation have demonstrated efficacy in feed-forward sparse-view reconstruction. However, such representation necessitates cross-view overlap for accurate depth estimation, and is challenged by object occlusions and frustum truncations. As a result, these methods require scene-centric data acquisition to maintain cross-view overlap and complete scene visibility to circumvent occlusions and truncations, which limits their applicability to scene-centric reconstruction. In contrast, in autonomous driving scenarios, a more practical paradigm is ego-centric reconstruction, which is characterized by minimal cross-view overlap and frequent occlusions and truncations. The limitations of pixel-based representation thus hinder the utility of prior works in this task. In light of this, this paper conducts an in-depth analysis of different representations, and introduces Omni-Gaussian representation with tailored network design to complement their strengths and mitigate their drawbacks. Experiments show that our method significantly surpasses state-of-the-art methods, pixelSplat and MVSplat, in ego-centric reconstruction, and achieves comparable performance to prior works in scene-centric reconstruction.

Jindong Jiang, Xiuyu Li, Zhijian Liu et al.

Recent advances in video-based multimodal large language models (Video-LLMs) have significantly improved video understanding by processing videos as sequences of image frames. However, many existing methods treat frames independently in the vision backbone, lacking explicit temporal modeling, which limits their ability to capture dynamic patterns and efficiently handle long videos. To address these limitations, we introduce STORM (Spatiotemporal TOken Reduction for Multimodal LLMs), a novel architecture incorporating a dedicated temporal encoder between the image encoder and the LLM. Our temporal encoder leverages the Mamba State Space Model to integrate temporal information into image tokens, generating enriched representations that preserve inter-frame dynamics across the entire video sequence. This enriched encoding not only enhances video reasoning capabilities but also enables effective token reduction strategies, including test-time sampling and training-based temporal and spatial pooling, substantially reducing computational demands on the LLM without sacrificing key temporal information. By integrating these techniques, our approach simultaneously reduces training and inference latency while improving performance, enabling efficient and robust video understanding over extended temporal contexts. Extensive evaluations show that STORM achieves state-of-the-art results across various long video understanding benchmarks (more than 5% improvement on MLVU and LongVideoBench) while reducing the computation costs by up to $8\times$ and the decoding latency by 2.4-2.9$\times$ for the fixed numbers of input frames. Project page is available at https://research.nvidia.com/labs/lpr/storm

Guo Chen, Lidong Lu, Yicheng Liu et al.

While datasets for video understanding have scaled to hour-long durations, they typically consist of densely concatenated clips that differ from natural, unscripted daily life. To bridge this gap, we introduce MM-Lifelong, a dataset designed for Multimodal Lifelong Understanding. Comprising 181.1 hours of footage, it is structured across Day, Week, and Month scales to capture varying temporal densities. Extensive evaluations reveal two critical failure modes in current paradigms: end-to-end MLLMs suffer from a Working Memory Bottleneck due to context saturation, while representative agentic baselines experience Global Localization Collapse when navigating sparse, month-long timelines. To address this, we propose the Recursive Multimodal Agent (ReMA), which employs dynamic memory management to iteratively update a recursive belief state, significantly outperforming existing methods. Finally, we establish dataset splits designed to isolate temporal and domain biases, providing a rigorous foundation for future research in supervised learning and out-of-distribution generalization.

Shihao Wang, Guo Chen, De-an Huang et al.

Recent studies have revealed that selecting informative and relevant video frames can significantly improve the performance of Video Large Language Models (Video-LLMs). Current methods, such as reducing inter-frame redundancy, employing separate models for image-text relevance assessment, or utilizing temporal video grounding for event localization, substantially adopt unsupervised learning paradigms, whereas they struggle to address the complex scenarios in long video understanding. We propose Instructed Temporal Grounding for Videos (VideoITG), featuring customized frame sampling aligned with user instructions. The core of VideoITG is the VidThinker pipeline, an automated annotation framework that explicitly mimics the human annotation process. First, it generates detailed clip-level captions conditioned on the instruction; then, it retrieves relevant video segments through instruction-guided reasoning; finally, it performs fine-grained frame selection to pinpoint the most informative visual evidence. Leveraging VidThinker, we construct the VideoITG-40K dataset, containing 40K videos and 500K instructed temporal grounding annotations. We then design a plug-and-play VideoITG model, which takes advantage of visual language alignment and reasoning capabilities of Video-LLMs, for effective frame selection in a discriminative manner. Coupled with Video-LLMs, VideoITG achieves consistent performance improvements across multiple multimodal video understanding benchmarks, showing its superiority and great potentials for video understanding.

Lidong Lu, Guo Chen, Zhiqi Li et al.

Despite progress in video understanding, current MLLMs struggle with counting tasks. Existing benchmarks are limited by short videos, close-set queries, lack of clue annotations, and weak multimodal coverage. In this paper, we introduce CG-AV-Counting, a manually-annotated clue-grounded counting benchmark with 1,027 multimodal questions and 5,845 annotated clues over 497 long videos. It supports both black-box and white-box evaluation, serving as a comprehensive testbed for both end-to-end and reasoning-based counting. To explore ways to improve model's counting capability, we propose AV-Reasoner, a model trained with GRPO and curriculum learning to generalize counting ability from related tasks. AV-Reasoner achieves state-of-the-art results across multiple benchmarks, demonstrating the effectiveness of reinforcement learning. However, experiments show that on out-of-domain benchmarks, reasoning in the language space fails to bring performance gains. The code and benchmark have been released on https://av-reasoner.github.io.

Chunyuan Deng, Zhiqi Li, Roy Xie et al.

Large Language Models (LLMs) are thought to struggle with arithmetic learning due to the inherent differences between language modeling and numerical computation, but concrete evidence has been lacking. This work responds to this claim through a two-side experiment. We first investigate whether LLMs leverage partial products during arithmetic learning. We find that although LLMs can identify some partial products after learning, they fail to leverage them for arithmetic tasks, conversely. We then explore how LLMs approach arithmetic symbolically by breaking tasks into subgroups, hypothesizing that difficulties arise from subgroup complexity and selection. Our results show that when subgroup complexity is fixed, LLMs treat a collection of different arithmetic operations similarly. By analyzing position-level accuracy across different training sizes, we further observe that it follows a U-shaped pattern: LLMs quickly learn the easiest patterns at the first and last positions, while progressively learning the more difficult patterns in the middle positions. This suggests that LLMs select subgroup following an easy-to-hard paradigm during learning. Our work confirms that LLMs are pure symbolic learners in arithmetic tasks and underscores the importance of understanding them deeply through subgroup-level quantification.

Zexi Li, Zhiqi Li, Jie Lin et al.

Model fusion aims to integrate several deep neural network (DNN) models' knowledge into one by fusing parameters, and it has promising applications, such as improving the generalization of foundation models and parameter averaging in federated learning. However, models under different settings (data, hyperparameter, etc.) have diverse neuron permutations; in other words, from the perspective of loss landscape, they reside in different loss basins, thus hindering model fusion performances. To alleviate this issue, previous studies highlighted the role of permutation invariance and have developed methods to find correct network permutations for neuron alignment after training. Orthogonal to previous attempts, this paper studies training-time neuron alignment, improving model fusion without the need for post-matching. Training-time alignment is cheaper than post-alignment and is applicable in various model fusion scenarios. Starting from fundamental hypotheses and theorems, a simple yet lossless algorithm called TNA-PFN is introduced. TNA-PFN utilizes partially fixed neuron weights as anchors to reduce the potential of training-time permutations, and it is empirically validated in reducing the barriers of linear mode connectivity and multi-model fusion. It is also validated that TNA-PFN can improve the fusion of pretrained models under the setting of model soup (vision transformers) and ColD fusion (pretrained language models). Based on TNA-PFN, two federated learning methods, FedPFN and FedPNU, are proposed, showing the prospects of training-time neuron alignment. FedPFN and FedPNU reach state-of-the-art performances in federated learning under heterogeneous settings and can be compatible with the server-side algorithm.

Yishen Ji, Ziyue Zhu, Zhenxin Zhu et al.

Recent progress in driving video generation has shown significant potential for enhancing self-driving systems by providing scalable and controllable training data. Although pretrained state-of-the-art generation models, guided by 2D layout conditions (e.g., HD maps and bounding boxes), can produce photorealistic driving videos, achieving controllable multi-view videos with high 3D consistency remains a major challenge. To tackle this, we introduce a novel spatial adaptive generation framework, CoGen, which leverages advances in 3D generation to improve performance in two key aspects: (i) To ensure 3D consistency, we first generate high-quality, controllable 3D conditions that capture the geometry of driving scenes. By replacing coarse 2D conditions with these fine-grained 3D representations, our approach significantly enhances the spatial consistency of the generated videos. (ii) Additionally, we introduce a consistency adapter module to strengthen the robustness of the model to multi-condition control. The results demonstrate that this method excels in preserving geometric fidelity and visual realism, offering a reliable video generation solution for autonomous driving.

Zhiqi Li, Chengrui Dong, Yiming Chen et al.

We present VicaSplat, a novel framework for joint 3D Gaussians reconstruction and camera pose estimation from a sequence of unposed video frames, which is a critical yet underexplored task in real-world 3D applications. The core of our method lies in a novel transformer-based network architecture. In particular, our model starts with an image encoder that maps each image to a list of visual tokens. All visual tokens are concatenated with additional inserted learnable camera tokens. The obtained tokens then fully communicate with each other within a tailored transformer decoder. The camera tokens causally aggregate features from visual tokens of different views, and further modulate them frame-wisely to inject view-dependent features. 3D Gaussian splats and camera pose parameters can then be estimated via different prediction heads. Experiments show that VicaSplat surpasses baseline methods for multi-view inputs, and achieves comparable performance to prior two-view approaches. Remarkably, VicaSplat also demonstrates exceptional cross-dataset generalization capability on the ScanNet benchmark, achieving superior performance without any fine-tuning. Project page: https://lizhiqi49.github.io/VicaSplat.

Yishen Ji, Zhiqi Li, Tong Lu

Track Mapless demands models to process multi-view images and Standard-Definition (SD) maps, outputting lane and traffic element perceptions along with their topological relationships. We propose a novel architecture that integrates SD map priors to improve lane line and area detection performance. Inspired by TopoMLP, our model employs a two-stage structure: perception and reasoning. The downstream topology head uses the output from the upstream detection head, meaning accuracy improvements in detection significantly boost downstream performance.

Zhiqi Li, Yuchen Sun, Greg Turk et al.

We present Functional Mean Flow (FMF) as a one-step generative model defined in infinite-dimensional Hilbert space. FMF extends the one-step Mean Flow framework to functional domains by providing a theoretical formulation for Functional Flow Matching and a practical implementation for efficient training and sampling. We also introduce an $x_1$-prediction variant that improves stability over the original $u$-prediction form. The resulting framework is a practical one-step Flow Matching method applicable to a wide range of functional data generation tasks such as time series, images, PDEs, and 3D geometry.

Yunze Man, Shihao Wang, Guowen Zhang et al.

To act in the world, a model must name what it sees and know where it is in 3D. Today's vision-language models (VLMs) excel at open-ended 2D description and grounding, yet multi-object 3D detection remains largely missing from the VLM toolbox. We present LocateAnything3D, a VLM-native recipe that casts 3D detection as a next-token prediction problem. The key is a short, explicit Chain-of-Sight (CoS) sequence that mirrors how human reason from images: find an object in 2D, then infer its distance, size, and pose. The decoder first emits 2D detections as a visual chain-of-thought, then predicts 3D boxes under an easy-to-hard curriculum: across objects, a near-to-far order reduces early ambiguity and matches ego-centric utility; within each object, a center-from-camera, dimensions, and rotation factorization ranks information by stability and learnability. This VLM-native interface preserves open-vocabulary and visual-prompting capability without specialized heads. On the challenging Omni3D benchmark, our model achieves state-of-the-art results, with 38.90 AP_3D, surpassing the previous best by +13.98 absolute improvement even when the baseline is given ground-truth 2D boxes. It also generalizes zero-shot to held-out categories with strong robustness. By turning 3D detection into a disciplined next-token problem, LocateAnything3D offers a practical foundation for models to perceive in 3D.