Xinlei Chen

Yu Sun, Xinhao Li, Karan Dalal et al.

Self-attention performs well in long context but has quadratic complexity. Existing RNN layers have linear complexity, but their performance in long context is limited by the expressive power of their hidden states. We present a practical framework for instantiating sequence modeling layers with linear complexity and expressive hidden states. The key idea is to make the hidden state a machine learning model itself, and the update rule a step of self-supervised learning. Since the hidden state is updated by training even on test sequences, our layers are called Test-Time Training (TTT) layers. We consider two instantiations: TTT-Linear and TTT-MLP, whose hidden state is a linear model and a two-layer MLP respectively. We evaluate our instantiations at the scale of 125M to 1.3B parameters, comparing with a strong Transformer and Mamba, a modern RNN. Similar to Transformer, TTT-Linear and TTT-MLP can keep reducing perplexity by conditioning on more tokens, while Mamba cannot after 16k context. TTT-MLP still faces challenges in memory I/O, but shows larger potential in long context, pointing to a promising direction for future research.

Lirui Wang, Xinlei Chen, Jialiang Zhao et al.

One of the roadblocks for training generalist robotic models today is heterogeneity. Previous robot learning methods often collect data to train with one specific embodiment for one task, which is expensive and prone to overfitting. This work studies the problem of learning policy representations through heterogeneous pre-training on robot data across different embodiments and tasks at scale. We propose Heterogeneous Pre-trained Transformers (HPT), which pre-train a large, shareable trunk of a policy neural network to learn a task and embodiment agnostic shared representation. This general architecture aligns the specific proprioception and vision inputs from distinct embodiments to a short sequence of tokens and then processes such tokens to map to control robots for different tasks. Leveraging the recent large-scale multi-embodiment real-world robotic datasets as well as simulation, deployed robots, and human video datasets, we investigate pre-training policies across heterogeneity. We conduct experiments to investigate the scaling behaviors of training objectives, to the extent of 52 datasets. HPTs outperform several baselines and enhance the fine-tuned policy performance by over 20% on unseen tasks in multiple simulator benchmarks and real-world settings. See the project website (https://liruiw.github.io/hpt/) for code and videos.

Nan Zhou, Huandong Wang, Jiahao Li et al.

Fine-grained fire prediction plays a crucial role in emergency response. Infrared images and fire masks provide complementary thermal and boundary information, yet current methods are predominantly limited to binary mask modeling with inherent signal sparsity, failing to capture the complex dynamics of fire. While world models show promise in video generation, their physical inconsistencies pose significant challenges for fire forecasting. This paper introduces PhysFire-WM, a Physics-informed World Model for emulating Fire spread dynamics. Our approach internalizes combustion dynamics by encoding structured priors from a Physical Simulator to rectify physical discrepancies, coupled with a Cross-task Collaborative Training strategy (CC-Train) that alleviates the issue of limited information in mask-based modeling. Through parameter sharing and gradient coordination, CC-Train effectively integrates thermal radiation dynamics and spatial boundary delineation, enhancing both physical realism and geometric accuracy. Extensive experiments on a fine-grained multimodal fire dataset demonstrate the superior accuracy of PhysFire-WM in fire spread prediction. Validation underscores the importance of physical priors and cross-task collaboration, providing new insights for applying physics-informed world models to disaster prediction.

Weichen Zhang, Chen Gao, Shiquan Yu et al.

Aerial vision-and-language navigation (VLN), requiring drones to interpret natural language instructions and navigate complex urban environments, emerges as a critical embodied AI challenge that bridges human-robot interaction, 3D spatial reasoning, and real-world deployment. Although existing ground VLN agents achieved notable results in indoor and outdoor settings, they struggle in aerial VLN due to the absence of predefined navigation graphs and the exponentially expanding action space in long-horizon exploration. In this work, we propose \textbf{CityNavAgent}, a large language model (LLM)-empowered agent that significantly reduces the navigation complexity for urban aerial VLN. Specifically, we design a hierarchical semantic planning module (HSPM) that decomposes the long-horizon task into sub-goals with different semantic levels. The agent reaches the target progressively by achieving sub-goals with different capacities of the LLM. Additionally, a global memory module storing historical trajectories into a topological graph is developed to simplify navigation for visited targets. Extensive benchmark experiments show that our method achieves state-of-the-art performance with significant improvement. Further experiments demonstrate the effectiveness of different modules of CityNavAgent for aerial VLN in continuous city environments. The code is available at \href{https://github.com/VinceOuti/CityNavAgent}{link}.

Yuandong Tian, Lantao Yu, Xinlei Chen et al.

We propose a novel theoretical framework to understand contrastive self-supervised learning (SSL) methods that employ dual pairs of deep ReLU networks (e.g., SimCLR). First, we prove that in each SGD update of SimCLR with various loss functions, including simple contrastive loss, soft Triplet loss and InfoNCE loss, the weights at each layer are updated by a \emph{covariance operator} that specifically amplifies initial random selectivities that vary across data samples but survive averages over data augmentations. To further study what role the covariance operator plays and which features are learned in such a process, we model data generation and augmentation processes through a \emph{hierarchical latent tree model} (HLTM) and prove that the hidden neurons of deep ReLU networks can learn the latent variables in HLTM, despite the fact that the network receives \emph{no direct supervision} from these unobserved latent variables. This leads to a provable emergence of hierarchical features through the amplification of initially random selectivities through contrastive SSL. Extensive numerical studies justify our theoretical findings. Code is released in https://github.com/facebookresearch/luckmatters/tree/master/ssl.

David Fan, Shengbang Tong, Jiachen Zhu et al.

Visual Self-Supervised Learning (SSL) currently underperforms Contrastive Language-Image Pretraining (CLIP) in multimodal settings such as Visual Question Answering (VQA). This multimodal gap is often attributed to the semantics introduced by language supervision, even though visual SSL and CLIP models are often trained on different data. In this work, we ask the question: "Do visual self-supervised approaches lag behind CLIP due to the lack of language supervision, or differences in the training data?" We study this question by training both visual SSL and CLIP models on the same MetaCLIP data, and leveraging VQA as a diverse testbed for vision encoders. In this controlled setup, visual SSL models scale better than CLIP models in terms of data and model capacity, and visual SSL performance does not saturate even after scaling up to 7B parameters. Consequently, we observe visual SSL methods achieve CLIP-level performance on a wide range of VQA and classic vision benchmarks. These findings demonstrate that pure visual SSL can match language-supervised visual pretraining at scale, opening new opportunities for vision-centric representation learning.

Jirong Zha, Yuxuan Fan, Kai Li et al.

State estimation is challenging for 3D object tracking with high maneuverability, as the target's state transition function changes rapidly, irregularly, and is unknown to the estimator. Existing work based on interacting multiple model (IMM) achieves more accurate estimation than single-filter approaches through model combination, aligning appropriate models for different motion modes of the target object over time. However, two limitations of conventional IMM remain unsolved. First, the solution space of the model combination is constrained as the target's diverse kinematic properties in different directions are ignored. Second, the model combination weights calculated by the observation likelihood are not accurate enough due to the measurement uncertainty. In this paper, we propose a novel framework, DIMM, to effectively combine estimates from different motion models in each direction, thus increasing the 3D object tracking accuracy. First, DIMM extends the model combination solution space of conventional IMM from a hyperplane to a hypercube by designing a 3D-decoupled multi-hierarchy filter bank, which describes the target's motion with various-order linear models. Second, DIMM generates more reliable combination weight matrices through a differentiable adaptive fusion network for importance allocation rather than solely relying on the observation likelihood; it contains an attention-based twin delayed deep deterministic policy gradient (TD3) method with a hierarchical reward. Experiments demonstrate that DIMM significantly improves the tracking accuracy of existing state estimation methods by 31.61%~99.23%.

Yu Meng, Kaiyuan Li, Chenran Huang et al.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across a range of multimodal tasks. However, their inference efficiency is constrained by the large number of visual tokens processed during decoding. To address this challenge, we propose Per-Layer Per-Head Vision Token Pruning (PLPHP), a two-level fine-grained pruning method including Layer-Level Retention Rate Allocation and Head-Level Vision Token Pruning. Motivated by the Vision Token Re-attention phenomenon across decoder layers, we dynamically adjust token retention rates layer by layer. Layers that exhibit stronger attention to visual information preserve more vision tokens, while layers with lower vision attention are aggressively pruned. Furthermore, PLPHP applies pruning at the attention head level, enabling different heads within the same layer to independently retain critical context. Experiments on multiple benchmarks demonstrate that PLPHP delivers an 18% faster decoding speed and reduces the Key-Value Cache (KV Cache) size by over 50%, all at the cost of 0.46% average performance drop, while also achieving notable performance improvements in multi-image tasks. These results highlight the effectiveness of fine-grained token pruning and contribute to advancing the efficiency and scalability of LVLMs. Our source code will be made publicly available.

Yanghao Li, Saining Xie, Xinlei Chen et al.

Object detection is a central downstream task used to test if pre-trained network parameters confer benefits, such as improved accuracy or training speed. The complexity of object detection methods can make this benchmarking non-trivial when new architectures, such as Vision Transformer (ViT) models, arrive. These difficulties (e.g., architectural incompatibility, slow training, high memory consumption, unknown training formulae, etc.) have prevented recent studies from benchmarking detection transfer learning with standard ViT models. In this paper, we present training techniques that overcome these challenges, enabling the use of standard ViT models as the backbone of Mask R-CNN. These tools facilitate the primary goal of our study: we compare five ViT initializations, including recent state-of-the-art self-supervised learning methods, supervised initialization, and a strong random initialization baseline. Our results show that recent masking-based unsupervised learning methods may, for the first time, provide convincing transfer learning improvements on COCO, increasing box AP up to 4% (absolute) over supervised and prior self-supervised pre-training methods. Moreover, these masking-based initializations scale better, with the improvement growing as model size increases.

Kaiming He, Xinlei Chen, Saining Xie et al.

This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. Our MAE approach is simple: we mask random patches of the input image and reconstruct the missing pixels. It is based on two core designs. First, we develop an asymmetric encoder-decoder architecture, with an encoder that operates only on the visible subset of patches (without mask tokens), along with a lightweight decoder that reconstructs the original image from the latent representation and mask tokens. Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task. Coupling these two designs enables us to train large models efficiently and effectively: we accelerate training (by 3x or more) and improve accuracy. Our scalable approach allows for learning high-capacity models that generalize well: e.g., a vanilla ViT-Huge model achieves the best accuracy (87.8%) among methods that use only ImageNet-1K data. Transfer performance in downstream tasks outperforms supervised pre-training and shows promising scaling behavior.

Kenneth Marino, Xinlei Chen, Devi Parikh et al.

One of the most challenging question types in VQA is when answering the question requires outside knowledge not present in the image. In this work we study open-domain knowledge, the setting when the knowledge required to answer a question is not given/annotated, neither at training nor test time. We tap into two types of knowledge representations and reasoning. First, implicit knowledge which can be learned effectively from unsupervised language pre-training and supervised training data with transformer-based models. Second, explicit, symbolic knowledge encoded in knowledge bases. Our approach combines both - exploiting the powerful implicit reasoning of transformer models for answer prediction, and integrating symbolic representations from a knowledge graph, while never losing their explicit semantics to an implicit embedding. We combine diverse sources of knowledge to cover the wide variety of knowledge needed to solve knowledge-based questions. We show our approach, KRISP (Knowledge Reasoning with Implicit and Symbolic rePresentations), significantly outperforms state-of-the-art on OK-VQA, the largest available dataset for open-domain knowledge-based VQA. We show with extensive ablations that while our model successfully exploits implicit knowledge reasoning, the symbolic answer module which explicitly connects the knowledge graph to the answer vocabulary is critical to the performance of our method and generalizes to rare answers.

Yuyin Zhou, Zhe Li, Song Bai et al.

Accurate multi-organ abdominal CT segmentation is essential to many clinical applications such as computer-aided intervention. As data annotation requires massive human labor from experienced radiologists, it is common that training data are partially labeled, e.g., pancreas datasets only have the pancreas labeled while leaving the rest marked as background. However, these background labels can be misleading in multi-organ segmentation since the "background" usually contains some other organs of interest. To address the background ambiguity in these partially-labeled datasets, we propose Prior-aware Neural Network (PaNN) via explicitly incorporating anatomical priors on abdominal organ sizes, guiding the training process with domain-specific knowledge. More specifically, PaNN assumes that the average organ size distributions in the abdomen should approximate their empirical distributions, a prior statistics obtained from the fully-labeled dataset. As our training objective is difficult to be directly optimized using stochastic gradient descent [20], we propose to reformulate it in a min-max form and optimize it via the stochastic primal-dual gradient algorithm. PaNN achieves state-of-the-art performance on the MICCAI2015 challenge "Multi-Atlas Labeling Beyond the Cranial Vault", a competition on organ segmentation in the abdomen. We report an average Dice score of 84.97%, surpassing the prior art by a large margin of 3.27%.

Xinlei Chen, Li-Jia Li, Li Fei-Fei et al.

We present a novel framework for iterative visual reasoning. Our framework goes beyond current recognition systems that lack the capability to reason beyond stack of convolutions. The framework consists of two core modules: a local module that uses spatial memory to store previous beliefs with parallel updates; and a global graph-reasoning module. Our graph module has three components: a) a knowledge graph where we represent classes as nodes and build edges to encode different types of semantic relationships between them; b) a region graph of the current image where regions in the image are nodes and spatial relationships between these regions are edges; c) an assignment graph that assigns regions to classes. Both the local module and the global module roll-out iteratively and cross-feed predictions to each other to refine estimates. The final predictions are made by combining the best of both modules with an attention mechanism. We show strong performance over plain ConvNets, \eg achieving an $8.4\%$ absolute improvement on ADE measured by per-class average precision. Analysis also shows that the framework is resilient to missing regions for reasoning.

Xinlei Chen, Abhinav Gupta

Modeling instance-level context and object-object relationships is extremely challenging. It requires reasoning about bounding boxes of different classes, locations \etc. Above all, instance-level spatial reasoning inherently requires modeling conditional distributions on previous detections. Unfortunately, our current object detection systems do not have any {\bf memory} to remember what to condition on! The state-of-the-art object detectors still detect all object in parallel followed by non-maximal suppression (NMS). While memory has been used for tasks such as captioning, they mostly use image-level memory cells without capturing the spatial layout. On the other hand, modeling object-object relationships requires {\bf spatial} reasoning -- not only do we need a memory to store the spatial layout, but also a effective reasoning module to extract spatial patterns. This paper presents a conceptually simple yet powerful solution -- Spatial Memory Network (SMN), to model the instance-level context efficiently and effectively. Our spatial memory essentially assembles object instances back into a pseudo "image" representation that is easy to be fed into another ConvNet for object-object context reasoning. This leads to a new sequential reasoning architecture where image and memory are processed in parallel to obtain detections which update the memory again. We show our SMN direction is promising as it provides 2.2\% improvement over baseline Faster RCNN on the COCO dataset so far.

Xinlei Chen, Abhinav Gupta

We adapted the join-training scheme of Faster RCNN framework from Caffe to TensorFlow as a baseline implementation for object detection. Our code is made publicly available. This report documents the simplifications made to the original pipeline, with justifications from ablation analysis on both PASCAL VOC 2007 and COCO 2014. We further investigated the role of non-maximal suppression (NMS) in selecting regions-of-interest (RoIs) for region classification, and found that a biased sampling toward small regions helps performance and can achieve on-par mAP to NMS-based sampling when converged sufficiently.

Xinlei Chen, Hao Fang, Tsung-Yi Lin et al.

In this paper we describe the Microsoft COCO Caption dataset and evaluation server. When completed, the dataset will contain over one and a half million captions describing over 330,000 images. For the training and validation images, five independent human generated captions will be provided. To ensure consistency in evaluation of automatic caption generation algorithms, an evaluation server is used. The evaluation server receives candidate captions and scores them using several popular metrics, including BLEU, METEOR, ROUGE and CIDEr. Instructions for using the evaluation server are provided.

Xinlei Chen, C. Lawrence Zitnick

In this paper we explore the bi-directional mapping between images and their sentence-based descriptions. We propose learning this mapping using a recurrent neural network. Unlike previous approaches that map both sentences and images to a common embedding, we enable the generation of novel sentences given an image. Using the same model, we can also reconstruct the visual features associated with an image given its visual description. We use a novel recurrent visual memory that automatically learns to remember long-term visual concepts to aid in both sentence generation and visual feature reconstruction. We evaluate our approach on several tasks. These include sentence generation, sentence retrieval and image retrieval. State-of-the-art results are shown for the task of generating novel image descriptions. When compared to human generated captions, our automatically generated captions are preferred by humans over $19.8\%$ of the time. Results are better than or comparable to state-of-the-art results on the image and sentence retrieval tasks for methods using similar visual features.