Yu-Xiong Wang

Zhiqing Sun, Sheng Shen, Shengcao Cao et al. · berkeley, cmu

Large Multimodal Models (LMM) are built across modalities and the misalignment between two modalities can result in "hallucination", generating textual outputs that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the task of vision-language alignment, where human annotators are asked to compare two responses and pinpoint the more hallucinated one, and the vision-language model is trained to maximize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image-text pairs to improve the general capabilities of our model. To evaluate the proposed approach in real-world scenarios, we develop a new evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaVA-Bench dataset with the 94% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improvement by 60% on MMHAL-BENCH over other baselines. We opensource our code, model, data at https://llava-rlhf.github.io.

Zhiqiu Lin, Deepak Pathak, Yu-Xiong Wang et al. · cmu

Lifelong learners must recognize concept vocabularies that evolve over time. A common yet underexplored scenario is learning with class labels that continually refine/expand old classes. For example, humans learn to recognize ${\tt dog}$ before dog breeds. In practical settings, dataset $\textit{versioning}$ often introduces refinement to ontologies, such as autonomous vehicle benchmarks that refine a previous ${\tt vehicle}$ class into ${\tt school-bus}$ as autonomous operations expand to new cities. This paper formalizes a protocol for studying the problem of $\textit{Learning with Evolving Class Ontology}$ (LECO). LECO requires learning classifiers in distinct time periods (TPs); each TP introduces a new ontology of "fine" labels that refines old ontologies of "coarse" labels (e.g., dog breeds that refine the previous ${\tt dog}$). LECO explores such questions as whether to annotate new data or relabel the old, how to leverage coarse labels, and whether to finetune the previous TP's model or train from scratch. To answer these questions, we leverage insights from related problems such as class-incremental learning. We validate them under the LECO protocol through the lens of image classification (CIFAR and iNaturalist) and semantic segmentation (Mapillary). Our experiments lead to surprising conclusions; while the current status quo is to relabel existing datasets with new ontologies (such as COCO-to-LVIS or Mapillary1.2-to-2.0), LECO demonstrates that a far better strategy is to annotate $\textit{new}$ data with the new ontology. However, this produces an aggregate dataset with inconsistent old-vs-new labels, complicating learning. To address this challenge, we adopt methods from semi-supervised and partial-label learning. Such strategies can surprisingly be made near-optimal, approaching an "oracle" that learns on the aggregate dataset exhaustively labeled with the newest ontology.

Ziqi Pang, Jie Li, Pavel Tokmakov et al.

This work proposes an end-to-end multi-camera 3D multi-object tracking (MOT) framework. It emphasizes spatio-temporal continuity and integrates both past and future reasoning for tracked objects. Thus, we name it "Past-and-Future reasoning for Tracking" (PF-Track). Specifically, our method adapts the "tracking by attention" framework and represents tracked instances coherently over time with object queries. To explicitly use historical cues, our "Past Reasoning" module learns to refine the tracks and enhance the object features by cross-attending to queries from previous frames and other objects. The "Future Reasoning" module digests historical information and predicts robust future trajectories. In the case of long-term occlusions, our method maintains the object positions and enables re-association by integrating motion predictions. On the nuScenes dataset, our method improves AMOTA by a large margin and remarkably reduces ID-Switches by 90% compared to prior approaches, which is an order of magnitude less. The code and models are made available at https://github.com/TRI-ML/PF-Track.

Ziqi Pang, Deva Ramanan, Mengtian Li et al.

Trajectory forecasting is a widely-studied problem for autonomous navigation. However, existing benchmarks evaluate forecasting based on independent snapshots of trajectories, which are not representative of real-world applications that operate on a continuous stream of data. To bridge this gap, we introduce a benchmark that continuously queries future trajectories on streaming data and we refer to it as "streaming forecasting." Our benchmark inherently captures the disappearance and re-appearance of agents, presenting the emergent challenge of forecasting for occluded agents, which is a safety-critical problem yet overlooked by snapshot-based benchmarks. Moreover, forecasting in the context of continuous timestamps naturally asks for temporal coherence between predictions from adjacent timestamps. Based on this benchmark, we further provide solutions and analysis for streaming forecasting. We propose a plug-and-play meta-algorithm called "Predictive Streamer" that can adapt any snapshot-based forecaster into a streaming forecaster. Our algorithm estimates the states of occluded agents by propagating their positions with multi-modal trajectories, and leverages differentiable filters to ensure temporal consistency. Both occlusion reasoning and temporal coherence strategies significantly improve forecasting quality, resulting in 25% smaller endpoint errors for occluded agents and 10-20% smaller fluctuations of trajectories. Our work is intended to generate interest within the community by highlighting the importance of addressing motion forecasting in its intrinsic streaming setting. Code is available at https://github.com/ziqipang/StreamingForecasting.

Xinglong Sun, Jean Ponce, Yu-Xiong Wang

Depth completion, which aims to generate high-quality dense depth maps from sparse depth maps, has attracted increasing attention in recent years. Previous work usually employs RGB images as guidance, and introduces iterative spatial propagation to refine estimated coarse depth maps. However, most of the propagation refinement methods require several iterations and suffer from a fixed receptive field, which may contain irrelevant and useless information with very sparse input. In this paper, we address these two challenges simultaneously by revisiting the idea of deformable convolution. We propose an effective architecture that leverages deformable kernel convolution as a single-pass refinement module, and empirically demonstrate its superiority. To better understand the function of deformable convolution and exploit it for depth completion, we further systematically investigate a variety of representative strategies. Our study reveals that, different from prior work, deformable convolution needs to be applied on an estimated depth map with a relatively high density for better performance. We evaluate our model on the large-scale KITTI dataset and achieve state-of-the-art level performance in both accuracy and inference speed. Our code is available at https://github.com/AlexSunNik/ReDC.

Sirui Xu, Yu-Xiong Wang, Liang-Yan Gui

Predicting diverse human motions given a sequence of historical poses has received increasing attention. Despite rapid progress, existing work captures the multi-modal nature of human motions primarily through likelihood-based sampling, where the mode collapse has been widely observed. In this paper, we propose a simple yet effective approach that disentangles randomly sampled codes with a deterministic learnable component named anchors to promote sample precision and diversity. Anchors are further factorized into spatial anchors and temporal anchors, which provide attractively interpretable control over spatial-temporal disparity. In principle, our spatial-temporal anchor-based sampling (STARS) can be applied to different motion predictors. Here we propose an interaction-enhanced spatial-temporal graph convolutional network (IE-STGCN) that encodes prior knowledge of human motions (e.g., spatial locality), and incorporate the anchors into it. Extensive experiments demonstrate that our approach outperforms state of the art in both stochastic and deterministic prediction, suggesting it as a unified framework for modeling human motions. Our code and pretrained models are available at https://github.com/Sirui-Xu/STARS.

Neehar Peri, Mengtian Li, Benjamin Wilson et al. · gatech

LiDAR-based 3D detection plays a vital role in autonomous navigation. Surprisingly, although autonomous vehicles (AVs) must detect both near-field objects (for collision avoidance) and far-field objects (for longer-term planning), contemporary benchmarks focus only on near-field 3D detection. However, AVs must detect far-field objects for safe navigation. In this paper, we present an empirical analysis of far-field 3D detection using the long-range detection dataset Argoverse 2.0 to better understand the problem, and share the following insight: near-field LiDAR measurements are dense and optimally encoded by small voxels, while far-field measurements are sparse and are better encoded with large voxels. We exploit this observation to build a collection of range experts tuned for near-vs-far field detection, and propose simple techniques to efficiently ensemble models for long-range detection that improve efficiency by 33% and boost accuracy by 3.2% CDS.

Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman et al.

While language models (LMs) have shown potential across a range of decision-making tasks, their reliance on simple acting processes limits their broad deployment as autonomous agents. In this paper, we introduce Language Agent Tree Search (LATS) -- the first general framework that synergizes the capabilities of LMs in reasoning, acting, and planning. By leveraging the in-context learning ability of LMs, we integrate Monte Carlo Tree Search into LATS to enable LMs as agents, along with LM-powered value functions and self-reflections for proficient exploration and enhanced decision-making. A key feature of our approach is the incorporation of an environment for external feedback, which offers a more deliberate and adaptive problem-solving mechanism that surpasses the constraints of existing techniques. Our experimental evaluation across diverse domains, including programming, interactive question-answering (QA), web navigation, and math, validates the effectiveness and generality of LATS in decision-making while maintaining competitive or improved reasoning performance. Notably, LATS achieves state-of-the-art pass@1 accuracy (92.7%) for programming on HumanEval with GPT-4 and demonstrates gradient-free performance (average score of 75.9) comparable to gradient-based fine-tuning for web navigation on WebShop with GPT-3.5. Code can be found at https://github.com/lapisrocks/LanguageAgentTreeSearch

Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan et al.

In the real open world, data tends to follow long-tailed class distributions, motivating the well-studied long-tailed recognition (LTR) problem. Naive training produces models that are biased toward common classes in terms of higher accuracy. The key to addressing LTR is to balance various aspects including data distribution, training losses, and gradients in learning. We explore an orthogonal direction, weight balancing, motivated by the empirical observation that the naively trained classifier has "artificially" larger weights in norm for common classes (because there exists abundant data to train them, unlike the rare classes). We investigate three techniques to balance weights, L2-normalization, weight decay, and MaxNorm. We first point out that L2-normalization "perfectly" balances per-class weights to be unit norm, but such a hard constraint might prevent classes from learning better classifiers. In contrast, weight decay penalizes larger weights more heavily and so learns small balanced weights; the MaxNorm constraint encourages growing small weights within a norm ball but caps all the weights by the radius. Our extensive study shows that both help learn balanced weights and greatly improve the LTR accuracy. Surprisingly, weight decay, although underexplored in LTR, significantly improves over prior work. Therefore, we adopt a two-stage training paradigm and propose a simple approach to LTR: (1) learning features using the cross-entropy loss by tuning weight decay, and (2) learning classifiers using class-balanced loss by tuning weight decay and MaxNorm. Our approach achieves the state-of-the-art accuracy on five standard benchmarks, serving as a future baseline for long-tailed recognition.

Kai Yan, Alexander G. Schwing, Yu-xiong Wang

In real-world scenarios, arbitrary interactions with the environment can often be costly, and actions of expert demonstrations are not always available. To reduce the need for both, offline Learning from Observations (LfO) is extensively studied: the agent learns to solve a task given only expert states and task-agnostic non-expert state-action pairs. The state-of-the-art DIstribution Correction Estimation (DICE) methods, as exemplified by SMODICE, minimize the state occupancy divergence between the learner's and empirical expert policies. However, such methods are limited to either $f$-divergences (KL and $chi^2$) or Wasserstein distance with Rubinstein duality, the latter of which constrains the underlying distance metric crucial to the performance of Wasserstein-based solutions. To enable more flexible distance metrics, we propose Primal Wasserstein DICE (PW-DICE). It minimizes the primal Wasserstein distance between the learner and expert state occupancies and leverages a contrastively learned distance metric. Theoretically, our framework is a generalization of SMODICE, and is the first work that unifies $f$-divergence and Wasserstein minimization. Empirically, we find that PW-DICE improves upon several state-of-the-art methods. The code is available at https://github.com/KaiYan289/PW-DICE.

Yunze Man, Shuhong Zheng, Zhipeng Bao et al.

Complex 3D scene understanding has gained increasing attention, with scene encoding strategies playing a crucial role in this success. However, the optimal scene encoding strategies for various scenarios remain unclear, particularly compared to their image-based counterparts. To address this issue, we present a comprehensive study that probes various visual encoding models for 3D scene understanding, identifying the strengths and limitations of each model across different scenarios. Our evaluation spans seven vision foundation encoders, including image-based, video-based, and 3D foundation models. We evaluate these models in four tasks: Vision-Language Scene Reasoning, Visual Grounding, Segmentation, and Registration, each focusing on different aspects of scene understanding. Our evaluations yield key findings: DINOv2 demonstrates superior performance, video models excel in object-level tasks, diffusion models benefit geometric tasks, and language-pretrained models show unexpected limitations in language-related tasks. These insights challenge some conventional understandings, provide novel perspectives on leveraging visual foundation models, and highlight the need for more flexible encoder selection in future vision-language and scene-understanding tasks. Code: https://github.com/YunzeMan/Lexicon3D

Zhipeng Bao, Pavel Tokmakov, Allan Jabri et al.

This paper studies the problem of object discovery -- separating objects from the background without manual labels. Existing approaches utilize appearance cues, such as color, texture, and location, to group pixels into object-like regions. However, by relying on appearance alone, these methods fail to separate objects from the background in cluttered scenes. This is a fundamental limitation since the definition of an object is inherently ambiguous and context-dependent. To resolve this ambiguity, we choose to focus on dynamic objects -- entities that can move independently in the world. We then scale the recent auto-encoder based frameworks for unsupervised object discovery from toy synthetic images to complex real-world scenes. To this end, we simplify their architecture, and augment the resulting model with a weak learning signal from general motion segmentation algorithms. Our experiments demonstrate that, despite only capturing a small subset of the objects that move, this signal is enough to generalize to segment both moving and static instances of dynamic objects. We show that our model scales to a newly collected, photo-realistic synthetic dataset with street driving scenarios. Additionally, we leverage ground truth segmentation and flow annotations in this dataset for thorough ablation and evaluation. Finally, our experiments on the real-world KITTI benchmark demonstrate that the proposed approach outperforms both heuristic- and learning-based methods by capitalizing on motion cues.

Jun-Kun Chen, Yu-Xiong Wang

Being able to learn an effective semantic representation directly on raw point clouds has become a central topic in 3D understanding. Despite rapid progress, state-of-the-art encoders are restrictive to canonicalized point clouds, and have weaker than necessary performance when encountering geometric transformation distortions. To overcome this challenge, we propose PointTree, a general-purpose point cloud encoder that is robust to transformations based on relaxed K-D trees. Key to our approach is the design of the division rule in K-D trees by using principal component analysis (PCA). We use the structure of the relaxed K-D tree as our computational graph, and model the features as border descriptors which are merged with pointwise-maximum operation. In addition to this novel architecture design, we further improve the robustness by introducing pre-alignment -- a simple yet effective PCA-based normalization scheme. Our PointTree encoder combined with pre-alignment consistently outperforms state-of-the-art methods by large margins, for applications from object classification to semantic segmentation on various transformed versions of the widely-benchmarked datasets. Code and pre-trained models are available at https://github.com/immortalCO/PointTree.

Zhipeng Bao, Pavel Tokmakov, Yu-Xiong Wang et al.

Object discovery -- separating objects from the background without manual labels -- is a fundamental open challenge in computer vision. Previous methods struggle to go beyond clustering of low-level cues, whether handcrafted (e.g., color, texture) or learned (e.g., from auto-encoders). In this work, we augment the auto-encoder representation learning framework with two key components: motion-guidance and mid-level feature tokenization. Although both have been separately investigated, we introduce a new transformer decoder showing that their benefits can compound thanks to motion-guided vector quantization. We show that our architecture effectively leverages the synergy between motion and tokenization, improving upon the state of the art on both synthetic and real datasets. Our approach enables the emergence of interpretable object-specific mid-level features, demonstrating the benefits of motion-guidance (no labeling) and quantization (interpretability, memory efficiency).

Ziqi Pang, Ziyang Xie, Yunze Man et al.

This paper reveals that large language models (LLMs), despite being trained solely on textual data, are surprisingly strong encoders for purely visual tasks in the absence of language. Even more intriguingly, this can be achieved by a simple yet previously overlooked strategy -- employing a frozen transformer block from pre-trained LLMs as a constituent encoder layer to directly process visual tokens. Our work pushes the boundaries of leveraging LLMs for computer vision tasks, significantly departing from conventional practices that typically necessitate a multi-modal vision-language setup with associated language prompts, inputs, or outputs. We demonstrate that our approach consistently enhances performance across a diverse range of tasks, encompassing pure 2D and 3D visual recognition tasks (e.g., image and point cloud classification), temporal modeling tasks (e.g., action recognition), non-semantic tasks (e.g., motion forecasting), and multi-modal tasks (e.g., 2D/3D visual question answering and image-text retrieval). Such improvements are a general phenomenon, applicable to various types of LLMs (e.g., LLaMA and OPT) and different LLM transformer blocks. We additionally propose the information filtering hypothesis to explain the effectiveness of pre-trained LLMs in visual encoding -- the pre-trained LLM transformer blocks discern informative visual tokens and further amplify their effect. This hypothesis is empirically supported by the observation that the feature activation, after training with LLM transformer blocks, exhibits a stronger focus on relevant regions. We hope that our work inspires new perspectives on utilizing LLMs and deepening our understanding of their underlying mechanisms. Code is available at https://github.com/ziqipang/LM4VisualEncoding.

Mingtong Zhang, Shuhong Zheng, Zhipeng Bao et al.

Comprehensive 3D scene understanding, both geometrically and semantically, is important for real-world applications such as robot perception. Most of the existing work has focused on developing data-driven discriminative models for scene understanding. This paper provides a new approach to scene understanding, from a synthesis model perspective, by leveraging the recent progress on implicit 3D representation and neural rendering. Building upon the great success of Neural Radiance Fields (NeRFs), we introduce Scene-Property Synthesis with NeRF (SS-NeRF) that is able to not only render photo-realistic RGB images from novel viewpoints, but also render various accurate scene properties (e.g., appearance, geometry, and semantics). By doing so, we facilitate addressing a variety of scene understanding tasks under a unified framework, including semantic segmentation, surface normal estimation, reshading, keypoint detection, and edge detection. Our SS-NeRF framework can be a powerful tool for bridging generative learning and discriminative learning, and thus be beneficial to the investigation of a wide range of interesting problems, such as studying task relationships within a synthesis paradigm, transferring knowledge to novel tasks, facilitating downstream discriminative tasks as ways of data augmentation, and serving as auto-labeller for data creation.

Michal Shlapentokh-Rothman, Prachi Garg, Yu-Xiong Wang et al.

Keyframe selection is a direct way to provide verifiable visual evidence for long-video question answering (QA). Queries differ in what they require, and finding the right frames depends on knowing what to look for. Existing keyframe selectors either score every frame against a single query, or decompose the query into a fixed schema evaluated by a single visual tool. We propose ToolMerge, a keyframe retrieval method based on decomposition and merging: an Large Language Model (LLM) based planner decomposes the query into tool calls and specifies how their per-tool rankings are merged using boolean operators. To evaluate retrieval directly, we construct Molmo-2 Moments (M2M), a benchmark in which every question is anchored to a specific time interval by construction. Across QA, question retrieval, and caption retrieval, ToolMerge is competitive with prior keyframe selectors, most notably on caption retrieval, outperforming other methods by 5%. Code and data can be found at https://github.com/michalsr/ToolMerge .

Yuanyi Zhong, Anand Bhattad, Yu-Xiong Wang et al.

Dense depth and surface normal predictors should possess the equivariant property to cropping-and-resizing -- cropping the input image should result in cropping the same output image. However, we find that state-of-the-art depth and normal predictors, despite having strong performances, surprisingly do not respect equivariance. The problem exists even when crop-and-resize data augmentation is employed during training. To remedy this, we propose an equivariant regularization technique, consisting of an averaging procedure and a self-consistency loss, to explicitly promote cropping-and-resizing equivariance in depth and normal networks. Our approach can be applied to both CNN and Transformer architectures, does not incur extra cost during testing, and notably improves the supervised and semi-supervised learning performance of dense predictors on Taskonomy tasks. Finally, finetuning with our loss on unlabeled images improves not only equivariance but also accuracy of state-of-the-art depth and normal predictors when evaluated on NYU-v2. GitHub link: https://github.com/mikuhatsune/equivariance

Sirui Xu, Zhengyuan Li, Yu-Xiong Wang et al.

This paper addresses a novel task of anticipating 3D human-object interactions (HOIs). Most existing research on HOI synthesis lacks comprehensive whole-body interactions with dynamic objects, e.g., often limited to manipulating small or static objects. Our task is significantly more challenging, as it requires modeling dynamic objects with various shapes, capturing whole-body motion, and ensuring physically valid interactions. To this end, we propose InterDiff, a framework comprising two key steps: (i) interaction diffusion, where we leverage a diffusion model to encode the distribution of future human-object interactions; (ii) interaction correction, where we introduce a physics-informed predictor to correct denoised HOIs in a diffusion step. Our key insight is to inject prior knowledge that the interactions under reference with respect to contact points follow a simple pattern and are easily predictable. Experiments on multiple human-object interaction datasets demonstrate the effectiveness of our method for this task, capable of producing realistic, vivid, and remarkably long-term 3D HOI predictions.

Kai Yan, Alexander G. Schwing, Yu-Xiong Wang

Offline imitation from observations aims to solve MDPs where only task-specific expert states and task-agnostic non-expert state-action pairs are available. Offline imitation is useful in real-world scenarios where arbitrary interactions are costly and expert actions are unavailable. The state-of-the-art "DIstribution Correction Estimation" (DICE) methods minimize divergence of state occupancy between expert and learner policies and retrieve a policy with weighted behavior cloning; however, their results are unstable when learning from incomplete trajectories, due to a non-robust optimization in the dual domain. To address the issue, in this paper, we propose Trajectory-Aware Imitation Learning from Observations (TAILO). TAILO uses a discounted sum along the future trajectory as the weight for weighted behavior cloning. The terms for the sum are scaled by the output of a discriminator, which aims to identify expert states. Despite simplicity, TAILO works well if there exist trajectories or segments of expert behavior in the task-agnostic data, a common assumption in prior work. In experiments across multiple testbeds, we find TAILO to be more robust and effective, particularly with incomplete trajectories.

Kuan-Ying Lee, Yuanyi Zhong, Yu-Xiong Wang

Existing work on continual learning (CL) is primarily devoted to developing algorithms for models trained from scratch. Despite their encouraging performance on contrived benchmarks, these algorithms show dramatic performance drops in real-world scenarios. Therefore, this paper advocates the systematic introduction of pre-training to CL, which is a general recipe for transferring knowledge to downstream tasks but is substantially missing in the CL community. Our investigation reveals the multifaceted complexity of exploiting pre-trained models for CL, along three different axes, pre-trained models, CL algorithms, and CL scenarios. Perhaps most intriguingly, improvements in CL algorithms from pre-training are very inconsistent an underperforming algorithm could become competitive and even state-of-the-art when all algorithms start from a pre-trained model. This indicates that the current paradigm, where all CL methods are compared in from-scratch training, is not well reflective of the true CL objective and desired progress. In addition, we make several other important observations, including that CL algorithms that exert less regularization benefit more from a pre-trained model; and that a stronger pre-trained model such as CLIP does not guarantee a better improvement. Based on these findings, we introduce a simple yet effective baseline that employs minimum regularization and leverages the more beneficial pre-trained model, coupled with a two-stage training pipeline. We recommend including this strong baseline in the future development of CL algorithms, due to its demonstrated state-of-the-art performance.

Shuhong Zheng, Zhipeng Bao, Martial Hebert et al.

Multi-task visual learning is a critical aspect of computer vision. Current research, however, predominantly concentrates on the multi-task dense prediction setting, which overlooks the intrinsic 3D world and its multi-view consistent structures, and lacks the capability for versatile imagination. In response to these limitations, we present a novel problem setting -- multi-task view synthesis (MTVS), which reinterprets multi-task prediction as a set of novel-view synthesis tasks for multiple scene properties, including RGB. To tackle the MTVS problem, we propose MuvieNeRF, a framework that incorporates both multi-task and cross-view knowledge to simultaneously synthesize multiple scene properties. MuvieNeRF integrates two key modules, the Cross-Task Attention (CTA) and Cross-View Attention (CVA) modules, enabling the efficient use of information across multiple views and tasks. Extensive evaluation on both synthetic and realistic benchmarks demonstrates that MuvieNeRF is capable of simultaneously synthesizing different scene properties with promising visual quality, even outperforming conventional discriminative models in various settings. Notably, we show that MuvieNeRF exhibits universal applicability across a range of NeRF backbones. Our code is available at https://github.com/zsh2000/MuvieNeRF.

Yuanyi Zhong, Haoran Tang, Junkun Chen et al.

Self-supervised contrastive learning is a powerful tool to learn visual representation without labels. Prior work has primarily focused on evaluating the recognition accuracy of various pre-training algorithms, but has overlooked other behavioral aspects. In addition to accuracy, distributional robustness plays a critical role in the reliability of machine learning models. We design and conduct a series of robustness tests to quantify the behavioral differences between contrastive learning and supervised learning to downstream or pre-training data distribution changes. These tests leverage data corruptions at multiple levels, ranging from pixel-level gamma distortion to patch-level shuffling and to dataset-level distribution shift. Our tests unveil intriguing robustness behaviors of contrastive and supervised learning. On the one hand, under downstream corruptions, we generally observe that contrastive learning is surprisingly more robust than supervised learning. On the other hand, under pre-training corruptions, we find contrastive learning vulnerable to patch shuffling and pixel intensity change, yet less sensitive to dataset-level distribution change. We attempt to explain these results through the role of data augmentation and feature space properties. Our insight has implications in improving the downstream robustness of supervised learning.

Sirui Xu, Yu-Xiong Wang, Liang-Yan Gui

This paper aims to deal with the ignored real-world complexities in prior work on human motion forecasting, emphasizing the social properties of multi-person motion, the diversity of motion and social interactions, and the complexity of articulated motion. To this end, we introduce a novel task of stochastic multi-person 3D motion forecasting. We propose a dual-level generative modeling framework that separately models independent individual motion at the local level and social interactions at the global level. Notably, this dual-level modeling mechanism can be achieved within a shared generative model, through introducing learnable latent codes that represent intents of future motion and switching the codes' modes of operation at different levels. Our framework is general; we instantiate it with different generative models, including generative adversarial networks and diffusion models, and various multi-person forecasting models. Extensive experiments on CMU-Mocap, MuPoTS-3D, and SoMoF benchmarks show that our approach produces diverse and accurate multi-person predictions, significantly outperforming the state of the art.

Kai Yan, Alexander G. Schwing, Yu-Xiong Wang

Although reinforcement learning has found widespread use in dense reward settings, training autonomous agents with sparse rewards remains challenging. To address this difficulty, prior work has shown promising results when using not only task-specific demonstrations but also task-agnostic albeit somewhat related demonstrations. In most cases, the available demonstrations are distilled into an implicit prior, commonly represented via a single deep net. Explicit priors in the form of a database that can be queried have also been shown to lead to encouraging results. To better benefit from available demonstrations, we develop a method to Combine Explicit and Implicit Priors (CEIP). CEIP exploits multiple implicit priors in the form of normalizing flows in parallel to form a single complex prior. Moreover, CEIP uses an effective explicit retrieval and push-forward mechanism to condition the implicit priors. In three challenging environments, we find the proposed CEIP method to improve upon sophisticated state-of-the-art techniques.

Brando Miranda, Patrick Yu, Yu-Xiong Wang et al.

Recently, it has been observed that a transfer learning solution might be all we need to solve many few-shot learning benchmarks -- thus raising important questions about when and how meta-learning algorithms should be deployed. In this paper, we seek to clarify these questions by 1. proposing a novel metric -- the diversity coefficient -- to measure the diversity of tasks in a few-shot learning benchmark and 2. by comparing Model-Agnostic Meta-Learning (MAML) and transfer learning under fair conditions (same architecture, same optimizer, and all models trained to convergence). Using the diversity coefficient, we show that the popular MiniImageNet and CIFAR-FS few-shot learning benchmarks have low diversity. This novel insight contextualizes claims that transfer learning solutions are better than meta-learned solutions in the regime of low diversity under a fair comparison. Specifically, we empirically find that a low diversity coefficient correlates with a high similarity between transfer learning and MAML learned solutions in terms of accuracy at meta-test time and classification layer similarity (using feature based distance metrics like SVCCA, PWCCA, CKA, and OPD). To further support our claim, we find this meta-test accuracy holds even as the model size changes. Therefore, we conclude that in the low diversity regime, MAML and transfer learning have equivalent meta-test performance when both are compared fairly. We also hope our work inspires more thoughtful constructions and quantitative evaluations of meta-learning benchmarks in the future.

Brando Miranda, Patrick Yu, Saumya Goyal et al.

In the context of few-shot learning, it is currently believed that a fixed pre-trained (PT) model, along with fine-tuning the final layer during evaluation, outperforms standard meta-learning algorithms. We re-evaluate these claims under an in-depth empirical examination of an extensive set of formally diverse datasets and compare PT to Model Agnostic Meta-Learning (MAML). Unlike previous work, we emphasize a fair comparison by using: the same architecture, the same optimizer, and all models trained to convergence. Crucially, we use a more rigorous statistical tool -- the effect size (Cohen's d) -- to determine the practical significance of the difference between a model trained with PT vs. a MAML. We then use a previously proposed metric -- the diversity coefficient -- to compute the average formal diversity of a dataset. Using this analysis, we demonstrate the following: 1. when the formal diversity of a data set is low, PT beats MAML on average and 2. when the formal diversity is high, MAML beats PT on average. The caveat is that the magnitude of the average difference between a PT vs. MAML using the effect size is low (according to classical statistical thresholds) -- less than 0.2. Nevertheless, this observation is contrary to the currently held belief that a pre-trained model is always better than a meta-learning model. Our extensive experiments consider 21 few-shot learning benchmarks, including the large-scale few-shot learning dataset Meta-Data set. We also show no significant difference between a MAML model vs. a PT model with GPT-2 on Openwebtext. We, therefore, conclude that a pre-trained model does not always beat a meta-learned model and that the formal diversity of a dataset is a driving factor.

Garvita Allabadi, Ana Lucic, Siddarth Aananth et al.

Traditional semi-supervised object detection methods assume a fixed set of object classes (in-distribution or ID classes) during training and deployment, which limits performance in real-world scenarios where unseen classes (out-of-distribution or OOD classes) may appear. In such cases, OOD data is often misclassified as ID, thus harming the ID classes accuracy. Open-set methods address this limitation by filtering OOD data to improve ID performance, thereby limiting the learning process to ID classes. We extend this to a more natural open-world setting, where the OOD classes are not only detected but also incorporated into the learning process. Specifically, we explore two key questions: 1) how to accurately detect OOD samples, and, most importantly, 2) how to effectively learn from the OOD samples in a semi-supervised object detection pipeline without compromising ID accuracy. To address this, we introduce an ensemble-based OOD Explorer for detection and classification, and an adaptable semi-supervised object detection framework that integrates both ID and OOD data. Through extensive evaluation on different open-world scenarios, we demonstrate that our method performs competitively against state-of-the-art OOD detection algorithms and also significantly boosts the semi-supervised learning performance for both ID and OOD classes.

Anurag Bagchi, Zhipeng Bao, Homanga Bharadhwaj et al.

What if a video generation model could not only imagine a plausible future, but the correct one, accurately reflecting how the world changes with each action? We address this question by presenting the Egocentric World Model (EgoWM), a simple, architecture-agnostic method that transforms any pretrained video diffusion model into an action-conditioned world model, enabling controllable future prediction. Rather than training from scratch, we repurpose the rich world priors of Internet-scale video models and inject motor commands through lightweight conditioning layers. This allows the model to follow actions faithfully while preserving realism and strong generalization. Our approach scales naturally across embodiments and action spaces, ranging from 3-DoF mobile robots to 25-DoF humanoids, where predicting egocentric joint-angle-driven dynamics is substantially more challenging. The model produces coherent rollouts for both navigation and manipulation tasks, requiring only modest fine-tuning. To evaluate physical correctness independently of visual appearance, we introduce the Structural Consistency Score (SCS), which measures whether stable scene elements evolve consistently with the provided actions. EgoWM improves SCS by up to 80 percent over prior state-of-the-art navigation world models, while achieving up to six times lower inference latency and robust generalization to unseen environments, including navigation inside paintings.

Yunze Man, Yichen Sheng, Jianming Zhang et al.

Recent advancements in 3D object reconstruction from single images have primarily focused on improving the accuracy of object shapes. Yet, these techniques often fail to accurately capture the inter-relation between the object, ground, and camera. As a result, the reconstructed objects often appear floating or tilted when placed on flat surfaces. This limitation significantly affects 3D-aware image editing applications like shadow rendering and object pose manipulation. To address this issue, we introduce ORG (Object Reconstruction with Ground), a novel task aimed at reconstructing 3D object geometry in conjunction with the ground surface. Our method uses two compact pixel-level representations to depict the relationship between camera, object, and ground. Experiments show that the proposed ORG model can effectively reconstruct object-ground geometry on unseen data, significantly enhancing the quality of shadow generation and pose manipulation compared to conventional single-image 3D reconstruction techniques.

Zheyu Zhang, Ziqi Pang, Shixing Chen et al.

Long video understanding is inherently challenging for vision-language models (VLMs) because of the extensive number of frames. With each video frame typically expanding into tens or hundreds of tokens, the limited context length of large language models (LLMs) forces the VLMs to perceive the frames sparsely and lose temporal information. To address this, we explore extreme video token compression towards one token per frame at the final LLM layer. Our key insight is that heuristic-based compression, widely adopted by previous methods, is prone to information loss, and this necessitates supervising LLM layers into learnable and progressive modules for token-level compression (LP-Comp). Such compression enables our VLM to digest 2x-4x more frames with improved performance. To further increase the token efficiency, we investigate frame-level compression, which selects the frames most relevant to the queries via the internal attention scores of the LLM layers, named question-conditioned compression (QC-Comp). As a notable distinction from previous studies, we mitigate the position bias of LLM attention in long contexts, i.e., the over-concentration on the beginning and end of a sequence, by splitting long videos into short segments and employing local attention. Collectively, our combined token-level and frame-level leads to an extreme compression model for long video understanding, named XComp, achieving a significantly larger compression ratio and enabling denser frame sampling. Our XComp is finetuned from VideoChat-Flash with a data-efficient supervised compression tuning stage that only requires 2.5% of the supervised fine-tuning data, yet boosts the accuracy from 42.9% to 46.2% on LVBench and enhances multiple other long video benchmarks.

Jiahua Dong, Yu-Xiong Wang

We introduce ViCA-NeRF, the first view-consistency-aware method for 3D editing with text instructions. In addition to the implicit neural radiance field (NeRF) modeling, our key insight is to exploit two sources of regularization that explicitly propagate the editing information across different views, thus ensuring multi-view consistency. For geometric regularization, we leverage the depth information derived from NeRF to establish image correspondences between different views. For learned regularization, we align the latent codes in the 2D diffusion model between edited and unedited images, enabling us to edit key views and propagate the update throughout the entire scene. Incorporating these two strategies, our ViCA-NeRF operates in two stages. In the initial stage, we blend edits from different views to create a preliminary 3D edit. This is followed by a second stage of NeRF training, dedicated to further refining the scene's appearance. Experimental results demonstrate that ViCA-NeRF provides more flexible, efficient (3 times faster) editing with higher levels of consistency and details, compared with the state of the art. Our code is publicly available.

Xialin He, Sirui Xu, Xinyao Li et al.

Achieving autonomous and versatile whole-body loco-manipulation remains a central barrier to making humanoids practically useful. Yet existing approaches are fundamentally constrained: retargeted data are often scarce or low-quality; methods struggle to scale to large skill repertoires; and, most importantly, they rely on tracking predefined motion references rather than generating behavior from perception and high-level task specifications. To address these limitations, we propose ULTRA, a unified framework with two key components. First, we introduce a physics-driven neural retargeting algorithm that translates large-scale motion capture to humanoid embodiments while preserving physical plausibility for contact-rich interactions. Second, we learn a unified multimodal controller that supports both dense references and sparse task specifications, under sensing ranging from accurate motion-capture state to noisy egocentric visual inputs. We distill a universal tracking policy into this controller, compress motor skills into a compact latent space, and apply reinforcement learning finetuning to expand coverage and improve robustness under out-of-distribution scenarios. This enables coordinated whole-body behavior from sparse intent without test-time reference motions. We evaluate ULTRA in simulation and on a real Unitree G1 humanoid. Results show that ULTRA generalizes to autonomous, goal-conditioned whole-body loco-manipulation from egocentric perception, consistently outperforming tracking-only baselines with limited skills.

Sirui Xu, Samuel Schulter, Morteza Ziyadi et al.

Humans rarely plan whole-body interactions with objects at the level of explicit whole-body movements. High-level intentions, such as affordance, define the goal, while coordinated balance, contact, and manipulation can emerge naturally from underlying physical and motor priors. Scaling such priors is key to enabling humanoids to compose and generalize loco-manipulation skills across diverse contexts while maintaining physically coherent whole-body coordination. To this end, we introduce InterPrior, a scalable framework that learns a unified generative controller through large-scale imitation pretraining and post-training by reinforcement learning. InterPrior first distills a full-reference imitation expert into a versatile, goal-conditioned variational policy that reconstructs motion from multimodal observations and high-level intent. While the distilled policy reconstructs training behaviors, it does not generalize reliably due to the vast configuration space of large-scale human-object interactions. To address this, we apply data augmentation with physical perturbations, and then perform reinforcement learning finetuning to improve competence on unseen goals and initializations. Together, these steps consolidate the reconstructed latent skills into a valid manifold, yielding a motion prior that generalizes beyond the training data, e.g., it can incorporate new behaviors such as interactions with unseen objects. We further demonstrate its effectiveness for user-interactive control and its potential for real robot deployment.

Andy Zhou, Jindong Wang, Yu-Xiong Wang et al.

We propose a conceptually simple and lightweight framework for improving the robustness of vision models through the combination of knowledge distillation and data augmentation. We address the conjecture that larger models do not make for better teachers by showing strong gains in out-of-distribution robustness when distilling from pretrained foundation models. Following this finding, we propose Discrete Adversarial Distillation (DAD), which leverages a robust teacher to generate adversarial examples and a VQGAN to discretize them, creating more informative samples than standard data augmentation techniques. We provide a theoretical framework for the use of a robust teacher in the knowledge distillation with data augmentation setting and demonstrate strong gains in out-of-distribution robustness and clean accuracy across different student architectures. Notably, our method adds minor computational overhead compared to similar techniques and can be easily combined with other data augmentations for further improvements.

Yifei Zhou, Qianlan Yang, Kaixiang Lin et al.

The vision of a broadly capable and goal-directed agent, such as an Internet-browsing agent in the digital world and a household humanoid in the physical world, has rapidly advanced, thanks to the generalization capability of foundation models. Such a generalist agent needs to have a large and diverse skill repertoire, such as finding directions between two travel locations and buying specific items from the Internet. If each skill needs to be specified manually through a fixed set of human-annotated instructions, the agent's skill repertoire will necessarily be limited due to the quantity and diversity of human-annotated instructions. In this work, we address this challenge by proposing Proposer-Agent-Evaluator, an effective learning system that enables foundation model agents to autonomously discover and practice skills in the wild. At the heart of PAE is a context-aware task proposer that autonomously proposes tasks for the agent to practice with context information of the environment such as user demos or even just the name of the website itself for Internet-browsing agents. Then, the agent policy attempts those tasks with thoughts and actual grounded operations in the real world with resulting trajectories evaluated by an autonomous VLM-based success evaluator. The success evaluation serves as the reward signal for the agent to refine its policies through RL. We validate PAE on challenging vision-based web navigation, using both real-world and self-hosted websites from WebVoyager and WebArena.To the best of our knowledge, this work represents the first effective learning system to apply autonomous task proposal with RL for agents that generalizes real-world human-annotated benchmarks with SOTA performances. Our open-source checkpoints and code can be found in https://yanqval.github.io/PAE/

Qiusi Zhan, Hyeonjeong Ha, Rui Yang et al.

Multimodal large language models (MLLMs) have advanced embodied agents by enabling direct perception, reasoning, and planning task-oriented actions from visual inputs. However, such vision driven embodied agents open a new attack surface: visual backdoor attacks, where the agent behaves normally until a visual trigger appears in the scene, then persistently executes an attacker-specified multi-step policy. We introduce BEAT, the first framework to inject such visual backdoors into MLLM-based embodied agents using objects in the environments as triggers. Unlike textual triggers, object triggers exhibit wide variation across viewpoints and lighting, making them difficult to implant reliably. BEAT addresses this challenge by (1) constructing a training set that spans diverse scenes, tasks, and trigger placements to expose agents to trigger variability, and (2) introducing a two-stage training scheme that first applies supervised fine-tuning (SFT) and then our novel Contrastive Trigger Learning (CTL). CTL formulates trigger discrimination as preference learning between trigger-present and trigger-free inputs, explicitly sharpening the decision boundaries to ensure precise backdoor activation. Across various embodied agent benchmarks and MLLMs, BEAT achieves attack success rates up to 80%, while maintaining strong benign task performance, and generalizes reliably to out-of-distribution trigger placements. Notably, compared to naive SFT, CTL boosts backdoor activation accuracy up to 39% under limited backdoor data. These findings expose a critical yet unexplored security risk in MLLM-based embodied agents, underscoring the need for robust defenses before real-world deployment.

Sirui Xu, Dongting Li, Yucheng Zhang et al.

While large-scale human motion capture datasets have advanced human motion generation, modeling and generating dynamic 3D human-object interactions (HOIs) remain challenging due to dataset limitations. Existing datasets often lack extensive, high-quality motion and annotation and exhibit artifacts such as contact penetration, floating, and incorrect hand motions. To address these issues, we introduce InterAct, a large-scale 3D HOI benchmark featuring dataset and methodological advancements. First, we consolidate and standardize 21.81 hours of HOI data from diverse sources, enriching it with detailed textual annotations. Second, we propose a unified optimization framework to enhance data quality by reducing artifacts and correcting hand motions. Leveraging the principle of contact invariance, we maintain human-object relationships while introducing motion variations, expanding the dataset to 30.70 hours. Third, we define six benchmarking tasks and develop a unified HOI generative modeling perspective, achieving state-of-the-art performance. Extensive experiments validate the utility of our dataset as a foundational resource for advancing 3D human-object interaction generation. To support continued research in this area, the dataset is publicly available at https://github.com/wzyabcas/InterAct, and will be actively maintained.

Michael Ofengenden, Yunze Man, Ziqi Pang et al.

We introduce PPTArena, a benchmark for PowerPoint editing that measures reliable modifications to real slides under natural-language instructions. In contrast to image-PDF renderings or text-to-slide generation, PPTArena focuses on in-place editing across 100 decks, 2125 slides, and over 800 targeted edits covering text, charts, tables, animations, and master-level styles. Each case includes a ground-truth deck, a fully specified target outcome, and a dual VLM-as-judge pipeline that separately scores instruction following and visual quality using both structural diffs and slide images. Building on this setting, we propose PPTPilot, a structure-aware slide-editing agent that plans semantic edit sequences, routes between high-level programmatic tools and deterministic XML operations for precise control, and verifies outputs through an iterative plan-edit-check loop against task-specific constraints. In our experiments, PPTPilot outperforms strong proprietary agents and frontier VLM systems by over 10 percentage points on compound, layout-sensitive, and cross-slide edits, with particularly large gains in visual fidelity and deck-wide consistency. Despite these improvements, existing agents still underperform on long-horizon, document-scale tasks in PPTArena, highlighting the remaining challenges in reliable PPT editing.

Ziyin Wang, Sirui Xu, Chuan Guo et al.

Generating realistic human-object interaction (HOI) animations remains challenging because it requires jointly modeling dynamic human actions and diverse object geometries. Prior diffusion-based approaches often rely on hand-crafted contact priors or human-imposed kinematic constraints to improve contact quality. We propose LIGHT, a data-driven alternative in which guidance emerges from the denoising pace itself, reducing dependence on manually designed priors. Building on diffusion forcing, we factor the representation into modality-specific components and assign individualized noise levels with asynchronous denoising schedules. In this paradigm, cleaner components guide noisier ones through cross-attention, yielding guidance without auxiliary classifiers. We find that this data-driven guidance is inherently contact-aware, and can be enhanced when training is augmented with a broad spectrum of synthetic object geometries, encouraging invariance of contact semantics to geometric diversity. Extensive experiments show that pace-induced guidance more effectively mirrors the benefits of contact priors than conventional classifier-free guidance, while achieving higher contact fidelity, more realistic HOI generation, and stronger generalization to unseen objects and tasks.

Ziqi Pang, Yu-Xiong Wang

We propose MR. Video, an agentic long video understanding framework that demonstrates the simple yet effective MapReduce principle for processing long videos: (1) Map: independently and densely perceiving short video clips, and (2) Reduce: jointly aggregating information from all clips. Compared with sequence-to-sequence vision-language models (VLMs), MR. Video performs detailed short video perception without being limited by context length. Compared with existing video agents that typically rely on sequential key segment selection, the Map operation enables simpler and more scalable sequence parallel perception of short video segments. Its Reduce step allows for more comprehensive context aggregation and reasoning, surpassing explicit key segment retrieval. This MapReduce principle is applicable to both VLMs and video agents, and we use LLM agents to validate its effectiveness. In practice, MR. Video employs two MapReduce stages: (A) Captioning: generating captions for short video clips (map), then standardizing repeated characters and objects into shared names (reduce); (B) Analysis: for each user question, analyzing relevant information from individual short videos (map), and integrating them into a final answer (reduce). MR. Video achieves over 10% accuracy improvement on the challenging LVBench compared to state-of-the-art VLMs and video agents. Code is available at: https://github.com/ziqipang/MR-Video

Zimu Zhang, Yucheng Zhang, Xiyan Xu et al.

Synthesizing human motion has advanced rapidly, yet realistic hand motion and bimanual interaction remain underexplored. Whole-body models often miss the fine-grained cues that drive dexterous behavior, finger articulation, contact timing, and inter-hand coordination, and existing resources lack high-fidelity bimanual sequences that capture nuanced finger dynamics and collaboration. To fill this gap, we present HandX, a unified foundation spanning data, annotation, and evaluation. We consolidate and filter existing datasets for quality, and collect a new motion-capture dataset targeting underrepresented bimanual interactions with detailed finger dynamics. For scalable annotation, we introduce a decoupled strategy that extracts representative motion features, e.g., contact events and finger flexion, and then leverages reasoning from large language models to produce fine-grained, semantically rich descriptions aligned with these features. Building on the resulting data and annotations, we benchmark diffusion and autoregressive models with versatile conditioning modes. Experiments demonstrate high-quality dexterous motion generation, supported by our newly proposed hand-focused metrics. We further observe clear scaling trends: larger models trained on larger, higher-quality datasets produce more semantically coherent bimanual motion. Our dataset is released to support future research.

Kai Yan, Alexander G. Schwing, Yu-Xiong Wang

Reinforcement Learning with Verifiable Rewards (RLVR) has achieved great success in developing Large Language Models (LLMs) with chain-of-thought rollouts for many tasks such as math and coding. Nevertheless, RLVR struggles with sample efficiency on difficult problems where correct rollouts are hard to generate. Prior works propose to address this issue via demonstration-guided RLVR, i.e., to conduct Supervised FineTuning (SFT) when RL fails; however, SFT often requires a lot of data, which can be expensive to acquire. In this paper, we propose FEST, a FEw-ShoT demonstration-guided RLVR algorithm. It attains compelling results with only 128 demonstrations randomly selected from an SFT dataset. We find that three components are vital for the success: supervised signal, on-policy signal, and decaying weights on the few-shot SFT dataset to prevent overfitting from multiple-epoch training. On several benchmarks, FEST outperforms baselines with magnitudes less SFT data, even matching their performance with full dataset.

Jun-Kun Chen, Aayush Bansal, Minh Phuoc Vo et al.

We present Dress&Dance, a video diffusion framework that generates high quality 5-second-long 24 FPS virtual try-on videos at 1152x720 resolution of a user wearing desired garments while moving in accordance with a given reference video. Our approach requires a single user image and supports a range of tops, bottoms, and one-piece garments, as well as simultaneous tops and bottoms try-on in a single pass. Key to our framework is CondNet, a novel conditioning network that leverages attention to unify multi-modal inputs (text, images, and videos), thereby enhancing garment registration and motion fidelity. CondNet is trained on heterogeneous training data, combining limited video data and a larger, more readily available image dataset, in a multistage progressive manner. Dress&Dance outperforms existing open source and commercial solutions and enables a high quality and flexible try-on experience.

Ziqi Pang, Xin Xu, Yu-Xiong Wang

With the success of image generation, generative diffusion models are increasingly adopted for discriminative tasks, as pixel generation provides a unified perception interface. However, directly repurposing the generative denoising process for discriminative objectives reveals critical gaps rarely addressed previously. Generative models tolerate intermediate sampling errors if the final distribution remains plausible, but discriminative tasks require rigorous accuracy throughout, as evidenced in challenging multi-modal tasks like referring image segmentation. Motivated by this gap, we analyze and enhance alignment between generative diffusion processes and perception tasks, focusing on how perception quality evolves during denoising. We find: (1) earlier denoising steps contribute disproportionately to perception quality, prompting us to propose tailored learning objectives reflecting varying timestep contributions; (2) later denoising steps show unexpected perception degradation, highlighting sensitivity to training-denoising distribution shifts, addressed by our diffusion-tailored data augmentation; and (3) generative processes uniquely enable interactivity, serving as controllable user interfaces adaptable to correctional prompts in multi-round interactions. Our insights significantly improve diffusion-based perception models without architectural changes, achieving state-of-the-art performance on depth estimation, referring image segmentation, and generalist perception tasks. Code available at https://github.com/ziqipang/ADDP.

Aruna Gauba, Irene Pi, Yunze Man et al.

We present AgMMU, a challenging real-world benchmark for evaluating and advancing vision-language models (VLMs) in the knowledge-intensive domain of agriculture. Unlike prior datasets that rely on crowdsourced prompts, AgMMU is distilled from 116,231 authentic dialogues between everyday growers and USDA-authorized Cooperative Extension experts. Through a three-stage pipeline: automated knowledge extraction, QA generation, and human verification, we construct (i) AgMMU, an evaluation set of 746 multiple-choice questions (MCQs) and 746 open-ended questions (OEQs), and (ii) AgBase, a development corpus of 57,079 multimodal facts covering five high-stakes agricultural topics: insect identification, species identification, disease categorization, symptom description, and management instruction. Benchmarking 12 leading VLMs reveals pronounced gaps in fine-grained perception and factual grounding. Open-sourced models trail after proprietary ones by a wide margin. Simple fine-tuning on AgBase boosts open-sourced model performance on challenging OEQs for up to 11.6% on average, narrowing this gap and also motivating future research to propose better strategies in knowledge extraction and distillation from AgBase. We hope AgMMU stimulates research on domain-specific knowledge integration and trustworthy decision support in agriculture AI development.

Saba Ghaffari, Ehsan Saleh, Alexander G. Schwing et al.

Protein design, a grand challenge of the day, involves optimization on a fitness landscape, and leading methods adopt a model-based approach where a model is trained on a training set (protein sequences and fitness) and proposes candidates to explore next. These methods are challenged by sparsity of high-fitness samples in the training set, a problem that has been in the literature. A less recognized but equally important problem stems from the distribution of training samples in the design space: leading methods are not designed for scenarios where the desired optimum is in a region that is not only poorly represented in training data, but also relatively far from the highly represented low-fitness regions. We show that this problem of "separation" in the design space is a significant bottleneck in existing model-based optimization tools and propose a new approach that uses a novel VAE as its search model to overcome the problem. We demonstrate its advantage over prior methods in robustly finding improved samples, regardless of the imbalance and separation between low- and high-fitness samples. Our comprehensive benchmark on real and semi-synthetic protein datasets as well as solution design for physics-informed neural networks, showcases the generality of our approach in discrete and continuous design spaces. Our implementation is available at https://github.com/sabagh1994/PGVAE.

Lue Fan, Ziqi Pang, Tianyuan Zhang et al.

In LiDAR-based 3D object detection for autonomous driving, the ratio of the object size to input scene size is significantly smaller compared to 2D detection cases. Overlooking this difference, many 3D detectors directly follow the common practice of 2D detectors, which downsample the feature maps even after quantizing the point clouds. In this paper, we start by rethinking how such multi-stride stereotype affects the LiDAR-based 3D object detectors. Our experiments point out that the downsampling operations bring few advantages, and lead to inevitable information loss. To remedy this issue, we propose Single-stride Sparse Transformer (SST) to maintain the original resolution from the beginning to the end of the network. Armed with transformers, our method addresses the problem of insufficient receptive field in single-stride architectures. It also cooperates well with the sparsity of point clouds and naturally avoids expensive computation. Eventually, our SST achieves state-of-the-art results on the large scale Waymo Open Dataset. It is worth mentioning that our method can achieve exciting performance (83.8 LEVEL 1 AP on validation split) on small object (pedestrian) detection due to the characteristic of single stride. Codes will be released at https://github.com/TuSimple/SST

Saba Ghaffari, Ehsan Saleh, David Forsyth et al.

Learning accurate classifiers for novel categories from very few examples, known as few-shot image classification, is a challenging task in statistical machine learning and computer vision. The performance in few-shot classification suffers from the bias in the estimation of classifier parameters; however, an effective underlying bias reduction technique that could alleviate this issue in training few-shot classifiers has been overlooked. In this work, we demonstrate the effectiveness of Firth bias reduction in few-shot classification. Theoretically, Firth bias reduction removes the $O(N^{-1})$ first order term from the small-sample bias of the Maximum Likelihood Estimator. Here we show that the general Firth bias reduction technique simplifies to encouraging uniform class assignment probabilities for multinomial logistic classification, and almost has the same effect in cosine classifiers. We derive an easy-to-implement optimization objective for Firth penalized multinomial logistic and cosine classifiers, which is equivalent to penalizing the cross-entropy loss with a KL-divergence between the uniform label distribution and the predictions. Then, we empirically evaluate that it is consistently effective across the board for few-shot image classification, regardless of (1) the feature representations from different backbones, (2) the number of samples per class, and (3) the number of classes. Finally, we show the robustness of Firth bias reduction, in the case of imbalanced data distribution. Our implementation is available at https://github.com/ehsansaleh/firth_bias_reduction

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.