Zhiding Yu

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.

Shikun Liu, Linxi Fan, Edward Johns et al. · stanford

Recent vision-language models have shown impressive multi-modal generation capabilities. However, typically they require training huge models on massive datasets. As a more scalable alternative, we introduce Prismer, a data- and parameter-efficient vision-language model that leverages an ensemble of task-specific experts. Prismer only requires training of a small number of components, with the majority of network weights inherited from multiple readily-available, pre-trained experts, and kept frozen during training. By leveraging experts from a wide range of domains, we show Prismer can efficiently pool this expert knowledge and adapt it to various vision-language reasoning tasks. In our experiments, we show that Prismer achieves fine-tuned and few-shot learning performance which is competitive with current state-of-the-arts, whilst requiring up to two orders of magnitude less training data. Code is available at https://github.com/NVlabs/prismer.

Yiming Li, Zhiding Yu, Christopher Choy et al.

Humans can easily imagine the complete 3D geometry of occluded objects and scenes. This appealing ability is vital for recognition and understanding. To enable such capability in AI systems, we propose VoxFormer, a Transformer-based semantic scene completion framework that can output complete 3D volumetric semantics from only 2D images. Our framework adopts a two-stage design where we start from a sparse set of visible and occupied voxel queries from depth estimation, followed by a densification stage that generates dense 3D voxels from the sparse ones. A key idea of this design is that the visual features on 2D images correspond only to the visible scene structures rather than the occluded or empty spaces. Therefore, starting with the featurization and prediction of the visible structures is more reliable. Once we obtain the set of sparse queries, we apply a masked autoencoder design to propagate the information to all the voxels by self-attention. Experiments on SemanticKITTI show that VoxFormer outperforms the state of the art with a relative improvement of 20.0% in geometry and 18.1% in semantics and reduces GPU memory during training to less than 16GB. Our code is available on https://github.com/NVlabs/VoxFormer.

Daquan Zhou, Zhiding Yu, Enze Xie et al.

Recent studies show that Vision Transformers(ViTs) exhibit strong robustness against various corruptions. Although this property is partly attributed to the self-attention mechanism, there is still a lack of systematic understanding. In this paper, we examine the role of self-attention in learning robust representations. Our study is motivated by the intriguing properties of the emerging visual grouping in Vision Transformers, which indicates that self-attention may promote robustness through improved mid-level representations. We further propose a family of fully attentional networks (FANs) that strengthen this capability by incorporating an attentional channel processing design. We validate the design comprehensively on various hierarchical backbones. Our model achieves a state-of-the-art 87.1% accuracy and 35.8% mCE on ImageNet-1k and ImageNet-C with 76.8M parameters. We also demonstrate state-of-the-art accuracy and robustness in two downstream tasks: semantic segmentation and object detection. Code is available at: https://github.com/NVlabs/FAN.

Yilun Chen, Zhiding Yu, Yukang Chen et al.

False negatives (FN) in 3D object detection, {\em e.g.}, missing predictions of pedestrians, vehicles, or other obstacles, can lead to potentially dangerous situations in autonomous driving. While being fatal, this issue is understudied in many current 3D detection methods. In this work, we propose Hard Instance Probing (HIP), a general pipeline that identifies \textit{FN} in a multi-stage manner and guides the models to focus on excavating difficult instances. For 3D object detection, we instantiate this method as FocalFormer3D, a simple yet effective detector that excels at excavating difficult objects and improving prediction recall. FocalFormer3D features a multi-stage query generation to discover hard objects and a box-level transformer decoder to efficiently distinguish objects from massive object candidates. Experimental results on the nuScenes and Waymo datasets validate the superior performance of FocalFormer3D. The advantage leads to strong performance on both detection and tracking, in both LiDAR and multi-modal settings. Notably, FocalFormer3D achieves a 70.5 mAP and 73.9 NDS on nuScenes detection benchmark, while the nuScenes tracking benchmark shows 72.1 AMOTA, both ranking 1st place on the nuScenes LiDAR leaderboard. Our code is available at \url{https://github.com/NVlabs/FocalFormer3D}.

Min Shi, Fuxiao Liu, Shihao Wang et al.

The ability to accurately interpret complex visual information is a crucial topic of multimodal large language models (MLLMs). Recent work indicates that enhanced visual perception significantly reduces hallucinations and improves performance on resolution-sensitive tasks, such as optical character recognition and document analysis. A number of recent MLLMs achieve this goal using a mixture of vision encoders. Despite their success, there is a lack of systematic comparisons and detailed ablation studies addressing critical aspects, such as expert selection and the integration of multiple vision experts. This study provides an extensive exploration of the design space for MLLMs using a mixture of vision encoders and resolutions. Our findings reveal several underlying principles common to various existing strategies, leading to a streamlined yet effective design approach. We discover that simply concatenating visual tokens from a set of complementary vision encoders is as effective as more complex mixing architectures or strategies. We additionally introduce Pre-Alignment to bridge the gap between vision-focused encoders and language tokens, enhancing model coherence. The resulting family of MLLMs, Eagle, surpasses other leading open-source models on major MLLM benchmarks.

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.

Zhuolin Yang, Wei Ping, Zihan Liu et al. · stanford

Augmenting pretrained language models (LMs) with a vision encoder (e.g., Flamingo) has obtained the state-of-the-art results in image-to-text generation. However, these models store all the knowledge within their parameters, thus often requiring enormous model parameters to model the abundant visual concepts and very rich textual descriptions. Additionally, they are inefficient in incorporating new data, requiring a computational-expensive fine-tuning process. In this work, we introduce a Retrieval-augmented Visual Language Model, Re-ViLM, built upon the Flamingo, that supports retrieving the relevant knowledge from the external database for zero and in-context few-shot image-to-text generations. By storing certain knowledge explicitly in the external database, our approach reduces the number of model parameters and can easily accommodate new data during evaluation by simply updating the database. We also construct an interleaved image and text data that facilitates in-context few-shot learning capabilities. We demonstrate that Re-ViLM significantly boosts performance for image-to-text generation tasks, especially for zero-shot and few-shot generation in out-of-domain settings with 4 times less parameters compared with baseline methods.

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.

De-An Huang, Zhiding Yu, Anima Anandkumar

We propose MinVIS, a minimal video instance segmentation (VIS) framework that achieves state-of-the-art VIS performance with neither video-based architectures nor training procedures. By only training a query-based image instance segmentation model, MinVIS outperforms the previous best result on the challenging Occluded VIS dataset by over 10% AP. Since MinVIS treats frames in training videos as independent images, we can drastically sub-sample the annotated frames in training videos without any modifications. With only 1% of labeled frames, MinVIS outperforms or is comparable to fully-supervised state-of-the-art approaches on YouTube-VIS 2019/2021. Our key observation is that queries trained to be discriminative between intra-frame object instances are temporally consistent and can be used to track instances without any manually designed heuristics. MinVIS thus has the following inference pipeline: we first apply the trained query-based image instance segmentation to video frames independently. The segmented instances are then tracked by bipartite matching of the corresponding queries. This inference is done in an online fashion and does not need to process the whole video at once. MinVIS thus has the practical advantages of reducing both the labeling costs and the memory requirements, while not sacrificing the VIS performance. Code is available at: https://github.com/NVlabs/MinVIS

Junfei Xiao, Zhichao Xu, Shiyi Lan et al.

This report describes the winning solution to the Robust Vision Challenge (RVC) semantic segmentation track at ECCV 2022. Our method adopts the FAN-B-Hybrid model as the encoder and uses SegFormer as the segmentation framework. The model is trained on a composite dataset consisting of images from 9 datasets (ADE20K, Cityscapes, Mapillary Vistas, ScanNet, VIPER, WildDash 2, IDD, BDD, and COCO) with a simple dataset balancing strategy. All the original labels are projected to a 256-class unified label space, and the model is trained using a cross-entropy loss. Without significant hyperparameter tuning or any specific loss weighting, our solution ranks the first place on all the testing semantic segmentation benchmarks from multiple domains (ADE20K, Cityscapes, Mapillary Vistas, ScanNet, VIPER, and WildDash 2). The proposed method can serve as a strong baseline for the multi-domain segmentation task and benefit future works. Code will be available at https://github.com/lambert-x/RVC_Segmentation.

Enze Xie, Zhiding Yu, Daquan Zhou et al.

In this paper, we propose M$^2$BEV, a unified framework that jointly performs 3D object detection and map segmentation in the Birds Eye View~(BEV) space with multi-camera image inputs. Unlike the majority of previous works which separately process detection and segmentation, M$^2$BEV infers both tasks with a unified model and improves efficiency. M$^2$BEV efficiently transforms multi-view 2D image features into the 3D BEV feature in ego-car coordinates. Such BEV representation is important as it enables different tasks to share a single encoder. Our framework further contains four important designs that benefit both accuracy and efficiency: (1) An efficient BEV encoder design that reduces the spatial dimension of a voxel feature map. (2) A dynamic box assignment strategy that uses learning-to-match to assign ground-truth 3D boxes with anchors. (3) A BEV centerness re-weighting that reinforces with larger weights for more distant predictions, and (4) Large-scale 2D detection pre-training and auxiliary supervision. We show that these designs significantly benefit the ill-posed camera-based 3D perception tasks where depth information is missing. M$^2$BEV is memory efficient, allowing significantly higher resolution images as input, with faster inference speed. Experiments on nuScenes show that M$^2$BEV achieves state-of-the-art results in both 3D object detection and BEV segmentation, with the best single model achieving 42.5 mAP and 57.0 mIoU in these two tasks, respectively.

Manli Shu, Weili Nie, De-An Huang et al.

Pre-trained vision-language models (e.g., CLIP) have shown promising zero-shot generalization in many downstream tasks with properly designed text prompts. Instead of relying on hand-engineered prompts, recent works learn prompts using the training data from downstream tasks. While effective, training on domain-specific data reduces a model's generalization capability to unseen new domains. In this work, we propose test-time prompt tuning (TPT), a method that can learn adaptive prompts on the fly with a single test sample. For image classification, TPT optimizes the prompt by minimizing the entropy with confidence selection so that the model has consistent predictions across different augmented views of each test sample. In evaluating generalization to natural distribution shifts, TPT improves the zero-shot top-1 accuracy of CLIP by 3.6% on average, surpassing previous prompt tuning approaches that require additional task-specific training data. In evaluating cross-dataset generalization with unseen categories, TPT performs on par with the state-of-the-art approaches that use additional training data. Project page: https://azshue.github.io/TPT.

Hao Zhang, Mingjie Liu, Shaokun Zhang et al.

Multi-turn LLM agents are increasingly important for solving complex, interactive tasks, and reinforcement learning (RL) is a key ingredient for improving their long-horizon behavior. However, RL training requires generating large numbers of sandboxed rollout trajectories, and existing infrastructures often couple rollout orchestration with the training loop, making systems hard to migrate and maintain. Under the rollout-as-a-service philosophy, we present ProRL Agent , a scalable infrastructure that serves the full agentic rollout lifecycle through an API service. ProRL Agent also provides standardized and extensible sandbox environments that support diverse agentic tasks in rootless HPC settings. We validate ProRL Agent through RL training on software engineering, math, STEM, and coding tasks. ProRL Agent is open-sourced and integrated as part of NVIDIA NeMo Gym.

Rafid Mahmood, James Lucas, David Acuna et al.

Given a small training data set and a learning algorithm, how much more data is necessary to reach a target validation or test performance? This question is of critical importance in applications such as autonomous driving or medical imaging where collecting data is expensive and time-consuming. Overestimating or underestimating data requirements incurs substantial costs that could be avoided with an adequate budget. Prior work on neural scaling laws suggest that the power-law function can fit the validation performance curve and extrapolate it to larger data set sizes. We find that this does not immediately translate to the more difficult downstream task of estimating the required data set size to meet a target performance. In this work, we consider a broad class of computer vision tasks and systematically investigate a family of functions that generalize the power-law function to allow for better estimation of data requirements. Finally, we show that incorporating a tuned correction factor and collecting over multiple rounds significantly improves the performance of the data estimators. Using our guidelines, practitioners can accurately estimate data requirements of machine learning systems to gain savings in both development time and data acquisition costs.

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.

Xiaojian Ma, Weili Nie, Zhiding Yu et al.

Reasoning about visual relationships is central to how humans interpret the visual world. This task remains challenging for current deep learning algorithms since it requires addressing three key technical problems jointly: 1) identifying object entities and their properties, 2) inferring semantic relations between pairs of entities, and 3) generalizing to novel object-relation combinations, i.e., systematic generalization. In this work, we use vision transformers (ViTs) as our base model for visual reasoning and make better use of concepts defined as object entities and their relations to improve the reasoning ability of ViTs. Specifically, we introduce a novel concept-feature dictionary to allow flexible image feature retrieval at training time with concept keys. This dictionary enables two new concept-guided auxiliary tasks: 1) a global task for promoting relational reasoning, and 2) a local task for facilitating semantic object-centric correspondence learning. To examine the systematic generalization of visual reasoning models, we introduce systematic splits for the standard HICO and GQA benchmarks. We show the resulting model, Concept-guided Vision Transformer (or RelViT for short) significantly outperforms prior approaches on HICO and GQA by 16% and 13% in the original split, and by 43% and 18% in the systematic split. Our ablation analyses also reveal our model's compatibility with multiple ViT variants and robustness to hyper-parameters.

Huaizu Jiang, Xiaojian Ma, Weili Nie et al.

A significant gap remains between today's visual pattern recognition models and human-level visual cognition especially when it comes to few-shot learning and compositional reasoning of novel concepts. We introduce Bongard-HOI, a new visual reasoning benchmark that focuses on compositional learning of human-object interactions (HOIs) from natural images. It is inspired by two desirable characteristics from the classical Bongard problems (BPs): 1) few-shot concept learning, and 2) context-dependent reasoning. We carefully curate the few-shot instances with hard negatives, where positive and negative images only disagree on action labels, making mere recognition of object categories insufficient to complete our benchmarks. We also design multiple test sets to systematically study the generalization of visual learning models, where we vary the overlap of the HOI concepts between the training and test sets of few-shot instances, from partial to no overlaps. Bongard-HOI presents a substantial challenge to today's visual recognition models. The state-of-the-art HOI detection model achieves only 62% accuracy on few-shot binary prediction while even amateur human testers on MTurk have 91% accuracy. With the Bongard-HOI benchmark, we hope to further advance research efforts in visual reasoning, especially in holistic perception-reasoning systems and better representation learning.

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.

Shiyi Lan, Xitong Yang, Zhiding Yu et al.

We propose Mask Auto-Labeler (MAL), a high-quality Transformer-based mask auto-labeling framework for instance segmentation using only box annotations. MAL takes box-cropped images as inputs and conditionally generates their mask pseudo-labels.We show that Vision Transformers are good mask auto-labelers. Our method significantly reduces the gap between auto-labeling and human annotation regarding mask quality. Instance segmentation models trained using the MAL-generated masks can nearly match the performance of their fully-supervised counterparts, retaining up to 97.4\% performance of fully supervised models. The best model achieves 44.1\% mAP on COCO instance segmentation (test-dev 2017), outperforming state-of-the-art box-supervised methods by significant margins. Qualitative results indicate that masks produced by MAL are, in some cases, even better than human annotations.

Jiachen Sun, Weili Nie, Zhiding Yu et al.

3D Point cloud is becoming a critical data representation in many real-world applications like autonomous driving, robotics, and medical imaging. Although the success of deep learning further accelerates the adoption of 3D point clouds in the physical world, deep learning is notorious for its vulnerability to adversarial attacks. In this work, we first identify that the state-of-the-art empirical defense, adversarial training, has a major limitation in applying to 3D point cloud models due to gradient obfuscation. We further propose PointDP, a purification strategy that leverages diffusion models to defend against 3D adversarial attacks. We extensively evaluate PointDP on six representative 3D point cloud architectures, and leverage 10+ strong and adaptive attacks to demonstrate its lower-bound robustness. Our evaluation shows that PointDP achieves significantly better robustness than state-of-the-art purification methods under strong attacks. Results of certified defenses on randomized smoothing combined with PointDP will be included in the near future.

Yiming Li, Sihang Li, Xinhao Liu et al.

Monocular scene understanding is a foundational component of autonomous systems. Within the spectrum of monocular perception topics, one crucial and useful task for holistic 3D scene understanding is semantic scene completion (SSC), which jointly completes semantic information and geometric details from RGB input. However, progress in SSC, particularly in large-scale street views, is hindered by the scarcity of high-quality datasets. To address this issue, we introduce SSCBench, a comprehensive benchmark that integrates scenes from widely used automotive datasets (e.g., KITTI-360, nuScenes, and Waymo). SSCBench follows an established setup and format in the community, facilitating the easy exploration of SSC methods in various street views. We benchmark models using monocular, trinocular, and point cloud input to assess the performance gap resulting from sensor coverage and modality. Moreover, we have unified semantic labels across diverse datasets to simplify cross-domain generalization testing. We commit to including more datasets and SSC models to drive further advancements in this field.

Zelun Luo, Yuliang Zou, Yijin Yang et al.

In recent years, differential privacy has seen significant advancements in image classification; however, its application to video activity recognition remains under-explored. This paper addresses the challenges of applying differential privacy to video activity recognition, which primarily stem from: (1) a discrepancy between the desired privacy level for entire videos and the nature of input data processed by contemporary video architectures, which are typically short, segmented clips; and (2) the complexity and sheer size of video datasets relative to those in image classification, which render traditional differential privacy methods inadequate. To tackle these issues, we propose Multi-Clip DP-SGD, a novel framework for enforcing video-level differential privacy through clip-based classification models. This method samples multiple clips from each video, averages their gradients, and applies gradient clipping in DP-SGD without incurring additional privacy loss. Moreover, we incorporate a parameter-efficient transfer learning strategy to make the model scalable for large-scale video datasets. Through extensive evaluations on the UCF-101 and HMDB-51 datasets, our approach exhibits impressive performance, achieving 81% accuracy with a privacy budget of epsilon=5 on UCF-101, marking a 76% improvement compared to a direct application of DP-SGD. Furthermore, we demonstrate that our transfer learning strategy is versatile and can enhance differentially private image classification across an array of datasets including CheXpert, ImageNet, CIFAR-10, and CIFAR-100.

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}

De-An Huang, Shijia Liao, Subhashree Radhakrishnan et al.

There has been tremendous progress in multimodal Large Language Models (LLMs). Recent works have extended these models to video input with promising instruction following capabilities. However, an important missing piece is temporal localization. These models cannot accurately answer the "When?" questions. We identify three key aspects that limit their temporal localization capabilities: (i) time representation, (ii) architecture, and (iii) data. We address these shortcomings by proposing Language Instructed Temporal-Localization Assistant (LITA) with the following features: (1) We introduce time tokens that encode timestamps relative to the video length to better represent time in videos. (2) We introduce SlowFast tokens in the architecture to capture temporal information at fine temporal resolution. (3) We emphasize temporal localization data for LITA. In addition to leveraging existing video datasets with timestamps, we propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning and evaluating this task. Reasoning temporal localization requires both the reasoning and temporal localization of Video LLMs. LITA demonstrates strong performance on this challenging task, nearly doubling the temporal mean intersection-over-union (mIoU) of baselines. In addition, we show that our emphasis on temporal localization also substantially improves video-based text generation compared to existing Video LLMs, including a 36% relative improvement of Temporal Understanding. Code is available at: https://github.com/NVlabs/LITA

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.

Zizheng Pan, Bohan Zhuang, De-An Huang et al.

Sampling from diffusion probabilistic models (DPMs) is often expensive for high-quality image generation and typically requires many steps with a large model. In this paper, we introduce sampling Trajectory Stitching T-Stitch, a simple yet efficient technique to improve the sampling efficiency with little or no generation degradation. Instead of solely using a large DPM for the entire sampling trajectory, T-Stitch first leverages a smaller DPM in the initial steps as a cheap drop-in replacement of the larger DPM and switches to the larger DPM at a later stage. Our key insight is that different diffusion models learn similar encodings under the same training data distribution and smaller models are capable of generating good global structures in the early steps. Extensive experiments demonstrate that T-Stitch is training-free, generally applicable for different architectures, and complements most existing fast sampling techniques with flexible speed and quality trade-offs. On DiT-XL, for example, 40% of the early timesteps can be safely replaced with a 10x faster DiT-S without performance drop on class-conditional ImageNet generation. We further show that our method can also be used as a drop-in technique to not only accelerate the popular pretrained stable diffusion (SD) models but also improve the prompt alignment of stylized SD models from the public model zoo. Code is released at https://github.com/NVlabs/T-Stitch

Barath Lakshmanan, Joshua Chen, Shiyi Lan et al.

The cornerstone of autonomous vehicles (AV) is a solid perception system, where camera encoders play a crucial role. Existing works usually leverage pre-trained Convolutional Neural Networks (CNN) or Vision Transformers (ViTs) designed for general vision tasks, such as image classification, segmentation, and 2D detection. Although those well-known architectures have achieved state-of-the-art accuracy in AV-related tasks, e.g., 3D Object Detection, there remains significant potential for improvement in network design due to the nuanced complexities of industrial-level AV dataset. Moreover, existing public AV benchmarks usually contain insufficient data, which might lead to inaccurate evaluation of those architectures.To reveal the AV-specific model insights, we start from a standard general-purpose encoder, ConvNeXt and progressively transform the design. We adjust different design parameters including width and depth of the model, stage compute ratio, attention mechanisms, and input resolution, supported by systematic analysis to each modifications. This customization yields an architecture optimized for AV camera encoder achieving 8.79% mAP improvement over the baseline. We believe our effort could become a sweet cookbook of image encoders for AV and pave the way to the next-level drive system.

De-An Huang, Subhashree Radhakrishnan, Zhiding Yu et al.

There has been impressive progress in Large Multimodal Models (LMMs). Recent works extend these models to long inputs, including multi-page documents and long videos. However, the model size and performance of these long context models are still limited due to the computational cost in both training and inference. In this work, we explore an orthogonal direction and process long inputs without long context LMMs. We propose Frame Selection Augmented Generation (FRAG), where the model first selects relevant frames within the input, and then only generates the final outputs based on the selected frames. The core of the selection process is done by scoring each frame independently, which does not require long context processing. The frames with the highest scores are then selected by a simple Top-K selection. We show that this frustratingly simple framework is applicable to both long videos and multi-page documents using existing LMMs without any fine-tuning. We consider two models, LLaVA-OneVision and InternVL2, in our experiments and show that FRAG consistently improves the performance and achieves state-of-the-art performances for both long video and long document understanding. For videos, FRAG substantially improves InternVL2-76B by 5.8% on MLVU and 3.7% on Video-MME. For documents, FRAG achieves over 20% improvements on MP-DocVQA compared with recent LMMs specialized in long document understanding. Code is available at: https://github.com/NVlabs/FRAG

Bingyin Zhao, Zhiding Yu, Shiyi Lan et al.

Recent studies indicate that Vision Transformers (ViTs) are robust against out-of-distribution scenarios. In particular, the Fully Attentional Network (FAN) - a family of ViT backbones, has achieved state-of-the-art robustness. In this paper, we revisit the FAN models and improve their pre-training with a self-emerging token labeling (STL) framework. Our method contains a two-stage training framework. Specifically, we first train a FAN token labeler (FAN-TL) to generate semantically meaningful patch token labels, followed by a FAN student model training stage that uses both the token labels and the original class label. With the proposed STL framework, our best model based on FAN-L-Hybrid (77.3M parameters) achieves 84.8% Top-1 accuracy and 42.1% mCE on ImageNet-1K and ImageNet-C, and sets a new state-of-the-art for ImageNet-A (46.1%) and ImageNet-R (56.6%) without using extra data, outperforming the original FAN counterpart by significant margins. The proposed framework also demonstrates significantly enhanced performance on downstream tasks such as semantic segmentation, with up to 1.7% improvement in robustness over the counterpart model. Code is available at https://github.com/NVlabs/STL.

Payman Behnam, Yaosheng Fu, Ritchie Zhao et al.

Transformer-based Large Language Models rely critically on the KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy containing two consecutive stages. In the first stage, it performs coarse-grain permanent KV cache eviction on the input sequence tokens. In the second stage, it adopts a hybrid sparse attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensionality reductions. We show that RocketKV provides a compression ratio of up to 400$\times$, end-to-end speedup of up to 3.7$\times$ as well as peak memory reduction of up to 32.6% in the decode phase on an NVIDIA A100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks. We also propose a variant of RocketKV for multi-turn scenarios, which consistently outperforms other existing methods and achieves accuracy nearly on par with an oracle top-k attention scheme. The source code is available here: https://github.com/NVlabs/RocketKV.

Junfei Xiao, Ziqi Zhou, Wenxuan Li et al.

This paper introduces ProLab, a novel approach using property-level label space for creating strong interpretable segmentation models. Instead of relying solely on category-specific annotations, ProLab uses descriptive properties grounded in common sense knowledge for supervising segmentation models. It is based on two core designs. First, we employ Large Language Models (LLMs) and carefully crafted prompts to generate descriptions of all involved categories that carry meaningful common sense knowledge and follow a structured format. Second, we introduce a description embedding model preserving semantic correlation across descriptions and then cluster them into a set of descriptive properties (e.g., 256) using K-Means. These properties are based on interpretable common sense knowledge consistent with theories of human recognition. We empirically show that our approach makes segmentation models perform stronger on five classic benchmarks (e.g., ADE20K, COCO-Stuff, Pascal Context, Cityscapes, and BDD). Our method also shows better scalability with extended training steps than category-level supervision. Our interpretable segmentation framework also emerges with the generalization ability to segment out-of-domain or unknown categories using only in-domain descriptive properties. Code is available at https://github.com/lambert-x/ProLab.

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}.

Xinlong Wang, Zhiding Yu, Shalini De Mello et al.

Instance segmentation is a fundamental vision task that aims to recognize and segment each object in an image. However, it requires costly annotations such as bounding boxes and segmentation masks for learning. In this work, we propose a fully unsupervised learning method that learns class-agnostic instance segmentation without any annotations. We present FreeSOLO, a self-supervised instance segmentation framework built on top of the simple instance segmentation method SOLO. Our method also presents a novel localization-aware pre-training framework, where objects can be discovered from complicated scenes in an unsupervised manner. FreeSOLO achieves 9.8% AP_{50} on the challenging COCO dataset, which even outperforms several segmentation proposal methods that use manual annotations. For the first time, we demonstrate unsupervised class-agnostic instance segmentation successfully. FreeSOLO's box localization significantly outperforms state-of-the-art unsupervised object detection/discovery methods, with about 100% relative improvements in COCO AP. FreeSOLO further demonstrates superiority as a strong pre-training method, outperforming state-of-the-art self-supervised pre-training methods by +9.8% AP when fine-tuning instance segmentation with only 5% COCO masks. Code is available at: github.com/NVlabs/FreeSOLO

Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura et al.

Deep neural networks on 3D point cloud data have been widely used in the real world, especially in safety-critical applications. However, their robustness against corruptions is less studied. In this paper, we present ModelNet40-C, the first comprehensive benchmark on 3D point cloud corruption robustness, consisting of 15 common and realistic corruptions. Our evaluation shows a significant gap between the performances on ModelNet40 and ModelNet40-C for state-of-the-art (SOTA) models. To reduce the gap, we propose a simple but effective method by combining PointCutMix-R and TENT after evaluating a wide range of augmentation and test-time adaptation strategies. We identify a number of critical insights for future studies on corruption robustness in point cloud recognition. For instance, we unveil that Transformer-based architectures with proper training recipes achieve the strongest robustness. We hope our in-depth analysis will motivate the development of robust training strategies or architecture designs in the 3D point cloud domain. Our codebase and dataset are included in https://github.com/jiachens/ModelNet40-C

Haotao Wang, Chaowei Xiao, Jean Kossaifi et al.

Data augmentation is a simple yet effective way to improve the robustness of deep neural networks (DNNs). Diversity and hardness are two complementary dimensions of data augmentation to achieve robustness. For example, AugMix explores random compositions of a diverse set of augmentations to enhance broader coverage, while adversarial training generates adversarially hard samples to spot the weakness. Motivated by this, we propose a data augmentation framework, termed AugMax, to unify the two aspects of diversity and hardness. AugMax first randomly samples multiple augmentation operators and then learns an adversarial mixture of the selected operators. Being a stronger form of data augmentation, AugMax leads to a significantly augmented input distribution which makes model training more challenging. To solve this problem, we further design a disentangled normalization module, termed DuBIN (Dual-Batch-and-Instance Normalization), that disentangles the instance-wise feature heterogeneity arising from AugMax. Experiments show that AugMax-DuBIN leads to significantly improved out-of-distribution robustness, outperforming prior arts by 3.03%, 3.49%, 1.82% and 0.71% on CIFAR10-C, CIFAR100-C, Tiny ImageNet-C and ImageNet-C. Codes and pretrained models are available: https://github.com/VITA-Group/AugMax.

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}.

Enze Xie, Wenhai Wang, Zhiding Yu et al.

We present SegFormer, a simple, efficient yet powerful semantic segmentation framework which unifies Transformers with lightweight multilayer perception (MLP) decoders. SegFormer has two appealing features: 1) SegFormer comprises a novel hierarchically structured Transformer encoder which outputs multiscale features. It does not need positional encoding, thereby avoiding the interpolation of positional codes which leads to decreased performance when the testing resolution differs from training. 2) SegFormer avoids complex decoders. The proposed MLP decoder aggregates information from different layers, and thus combining both local attention and global attention to render powerful representations. We show that this simple and lightweight design is the key to efficient segmentation on Transformers. We scale our approach up to obtain a series of models from SegFormer-B0 to SegFormer-B5, reaching significantly better performance and efficiency than previous counterparts. For example, SegFormer-B4 achieves 50.3% mIoU on ADE20K with 64M parameters, being 5x smaller and 2.2% better than the previous best method. Our best model, SegFormer-B5, achieves 84.0% mIoU on Cityscapes validation set and shows excellent zero-shot robustness on Cityscapes-C. Code will be released at: github.com/NVlabs/SegFormer.

Nadine Chang, Zhiding Yu, Yu-Xiong Wang et al.

Training on datasets with long-tailed distributions has been challenging for major recognition tasks such as classification and detection. To deal with this challenge, image resampling is typically introduced as a simple but effective approach. However, we observe that long-tailed detection differs from classification since multiple classes may be present in one image. As a result, image resampling alone is not enough to yield a sufficiently balanced distribution at the object level. We address object-level resampling by introducing an object-centric memory replay strategy based on dynamic, episodic memory banks. Our proposed strategy has two benefits: 1) convenient object-level resampling without significant extra computation, and 2) implicit feature-level augmentation from model updates. We show that image-level and object-level resamplings are both important, and thus unify them with a joint resampling strategy (RIO). Our method outperforms state-of-the-art long-tailed detection and segmentation methods on LVIS v0.5 across various backbones. Code is available at https://github.com/NVlabs/RIO.

Xueyan Zou, Fanyi Xiao, Zhiding Yu et al.

Aliasing refers to the phenomenon that high frequency signals degenerate into completely different ones after sampling. It arises as a problem in the context of deep learning as downsampling layers are widely adopted in deep architectures to reduce parameters and computation. The standard solution is to apply a low-pass filter (e.g., Gaussian blur) before downsampling. However, it can be suboptimal to apply the same filter across the entire content, as the frequency of feature maps can vary across both spatial locations and feature channels. To tackle this, we propose an adaptive content-aware low-pass filtering layer, which predicts separate filter weights for each spatial location and channel group of the input feature maps. We investigate the effectiveness and generalization of the proposed method across multiple tasks including ImageNet classification, COCO instance segmentation, and Cityscapes semantic segmentation. Qualitative and quantitative results demonstrate that our approach effectively adapts to the different feature frequencies to avoid aliasing while preserving useful information for recognition. Code is available at https://maureenzou.github.io/ddac/.

Wuyang Chen, Zhiding Yu, Zhangyang Wang et al.

Models trained on synthetic images often face degraded generalization to real data. As a convention, these models are often initialized with ImageNet pre-trained representation. Yet the role of ImageNet knowledge is seldom discussed despite common practices that leverage this knowledge to maintain the generalization ability. An example is the careful hand-tuning of early stopping and layer-wise learning rates, which is shown to improve synthetic-to-real generalization but is also laborious and heuristic. In this work, we explicitly encourage the synthetically trained model to maintain similar representations with the ImageNet pre-trained model, and propose a \textit{learning-to-optimize (L2O)} strategy to automate the selection of layer-wise learning rates. We demonstrate that the proposed framework can significantly improve the synthetic-to-real generalization performance without seeing and training on real data, while also benefiting downstream tasks such as domain adaptation. Code is available at: https://github.com/NVlabs/ASG.

Zhongzheng Ren, Zhiding Yu, Xiaodong Yang et al.

Weakly supervised learning has emerged as a compelling tool for object detection by reducing the need for strong supervision during training. However, major challenges remain: (1) differentiation of object instances can be ambiguous; (2) detectors tend to focus on discriminative parts rather than entire objects; (3) without ground truth, object proposals have to be redundant for high recalls, causing significant memory consumption. Addressing these challenges is difficult, as it often requires to eliminate uncertainties and trivial solutions. To target these issues we develop an instance-aware and context-focused unified framework. It employs an instance-aware self-training algorithm and a learnable Concrete DropBlock while devising a memory-efficient sequential batch back-propagation. Our proposed method achieves state-of-the-art results on COCO ($12.1\% ~AP$, $24.8\% ~AP_{50}$), VOC 2007 ($54.9\% ~AP$), and VOC 2012 ($52.1\% ~AP$), improving baselines by great margins. In addition, the proposed method is the first to benchmark ResNet based models and weakly supervised video object detection. Code, models, and more details will be made available at: https://github.com/NVlabs/wetectron.

Yang Zou, Zhiding Yu, Xiaofeng Liu et al.

Recent advances in domain adaptation show that deep self-training presents a powerful means for unsupervised domain adaptation. These methods often involve an iterative process of predicting on target domain and then taking the confident predictions as pseudo-labels for retraining. However, since pseudo-labels can be noisy, self-training can put overconfident label belief on wrong classes, leading to deviated solutions with propagated errors. To address the problem, we propose a confidence regularized self-training (CRST) framework, formulated as regularized self-training. Our method treats pseudo-labels as continuous latent variables jointly optimized via alternating optimization. We propose two types of confidence regularization: label regularization (LR) and model regularization (MR). CRST-LR generates soft pseudo-labels while CRST-MR encourages the smoothness on network output. Extensive experiments on image classification and semantic segmentation show that CRSTs outperform their non-regularized counterpart with state-of-the-art performance. The code and models of this work are available at https://github.com/yzou2/CRST.

Shihao Wang, Zhiding Yu, Xiaohui Jiang et al.

The advances in vision-language models (VLMs) have led to a growing interest in autonomous driving to leverage their strong reasoning capabilities. However, extending these capabilities from 2D to full 3D understanding is crucial for real-world applications. To address this challenge, we propose OmniDrive, a holistic vision-language dataset that aligns agent models with 3D driving tasks through counterfactual reasoning. This approach enhances decision-making by evaluating potential scenarios and their outcomes, similar to human drivers considering alternative actions. Our counterfactual-based synthetic data annotation process generates large-scale, high-quality datasets, providing denser supervision signals that bridge planning trajectories and language-based reasoning. Futher, we explore two advanced OmniDrive-Agent frameworks, namely Omni-L and Omni-Q, to assess the importance of vision-language alignment versus 3D perception, revealing critical insights into designing effective LLM-agents. Significant improvements on the DriveLM Q\&A benchmark and nuScenes open-loop planning demonstrate the effectiveness of our dataset and methods.

Shihao Wang, Zhiding Yu, Xiaohui Jiang et al.

The advances in vision-language models (VLMs) have led to a growing interest in autonomous driving to leverage their strong reasoning capabilities. However, extending these capabilities from 2D to full 3D understanding is crucial for real-world applications. To address this challenge, we propose OmniDrive, a holistic vision-language dataset that aligns agent models with 3D driving tasks through counterfactual reasoning. This approach enhances decision-making by evaluating potential scenarios and their outcomes, similar to human drivers considering alternative actions. Our counterfactual-based synthetic data annotation process generates large-scale, high-quality datasets, providing denser supervision signals that bridge planning trajectories and language-based reasoning. Futher, we explore two advanced OmniDrive-Agent frameworks, namely Omni-L and Omni-Q, to assess the importance of vision-language alignment versus 3D perception, revealing critical insights into designing effective LLM-agents. Significant improvements on the DriveLM Q\&A benchmark and nuScenes open-loop planning demonstrate the effectiveness of our dataset and methods.

Chonghao Sima, Kashyap Chitta, Zhiding Yu et al.

How can we rely on an end-to-end autonomous vehicle's complex decision-making system during deployment? One common solution is to have a ``fallback layer'' that checks the planned trajectory for rule violations and replaces it with a pre-defined safe action if necessary. Another approach involves adjusting the planner's decisions to minimize a pre-defined ``cost function'' using additional system predictions such as road layouts and detected obstacles. However, these pre-programmed rules or cost functions cannot learn and improve with new training data, often resulting in overly conservative behaviors. In this work, we propose Centaur (Cluster Entropy for Test-time trAining using Uncertainty) which updates a planner's behavior via test-time training, without relying on hand-engineered rules or cost functions. Instead, we measure and minimize the uncertainty in the planner's decisions. For this, we develop a novel uncertainty measure, called Cluster Entropy, which is simple, interpretable, and compatible with state-of-the-art planning algorithms. Using data collected at prior test-time time-steps, we perform an update to the model's parameters using a gradient that minimizes the Cluster Entropy. With only this sole gradient update prior to inference, Centaur exhibits significant improvements, ranking first on the navtest leaderboard with notable gains in safety-critical metrics such as time to collision. To provide detailed insights on a per-scenario basis, we also introduce navsafe, a challenging new benchmark, which highlights previously undiscovered failure modes of driving models.

Yue Zhao, Fuzhao Xue, Scott Reed et al.

We introduce Quantized Language-Image Pretraining (QLIP), a visual tokenization method that combines state-of-the-art reconstruction quality with state-of-the-art zero-shot image understanding. QLIP trains a binary-spherical-quantization-based autoencoder with reconstruction and language-image alignment objectives. We are the first to show that the two objectives do not need to be at odds. We balance the two loss terms dynamically during training and show that a two-stage training pipeline effectively mixes the large-batch requirements of image-language pre-training with the memory bottleneck imposed by the reconstruction objective. We validate the effectiveness of QLIP for multimodal understanding and text-conditioned image generation with a single model. Specifically, QLIP serves as a drop-in replacement for the visual encoder for LLaVA and the image tokenizer for LlamaGen with comparable or even better performance. Finally, we demonstrate that QLIP enables a unified mixed-modality auto-regressive model for understanding and generation.

Shaokun Zhang, Yi Dong, Jieyu Zhang et al.

Enabling large language models with external tools has become a pivotal strategy for extending their functionality beyond text space. To enhance LLMs' tool-calling abilities, previous approaches primarily rely on supervised fine-tuning (SFT) with trajectories distilled from stronger models, often resulting in imitative reasoning that limits generalization. In this work, we explore rule-based reinforcement learning to enhance tool-calling in LLMs, resulting in Nemotron-Research-Tool-N1, a series of tool-calling reasoning models. Rather than enforcing supervision over intermediate distilled reasoning traces, Tool-N1 is trained with a binary RL reward that assesses only the format validity and functional correctness of tool invocations. This lightweight supervision allows the model to develop reasoning strategies independently, without relying on annotated trajectories. Experiments on several major benchmarks show that Tool-N1-7B/14B clearly outperform GPT-4o. We conduct a systematic study on the design of rule-based reinforcement learning strategies for training tool-calling models. Using 5,518 distilled reasoning trajectories, we compare SFT, RL, and the SFT-then-RL pipeline, finding that the widely adopted SFT-then-RL paradigm does not necessarily outperform pure RL.

Mengyao Xu, Gabriel Moreira, Ronay Ak et al.

Motivated by the growing demand for retrieval systems that operate across modalities, we introduce llama-nemoretriever-colembed, a unified text-image retrieval model that delivers state-of-the-art performance across multiple benchmarks. We release two model variants, 1B and 3B. The 3B model achieves state of the art performance, scoring NDCG@5 91.0 on ViDoRe V1 and 63.5 on ViDoRe V2, placing first on both leaderboards as of June 27, 2025. Our approach leverages the NVIDIA Eagle2 Vision-Language model (VLM), modifies its architecture by replacing causal attention with bidirectional attention, and integrates a ColBERT-style late interaction mechanism to enable fine-grained multimodal retrieval in a shared embedding space. While this mechanism delivers superior retrieval accuracy, it introduces trade-offs in storage and efficiency. We provide a comprehensive analysis of these trade-offs. Additionally, we adopt a two-stage training strategy to enhance the model's retrieval capabilities.