Nicolas Ballas

Maxime Oquab, Timothée Darcet, Théo Moutakanni et al. · meta-ai, mit

The recent breakthroughs in natural language processing for model pretraining on large quantities of data have opened the way for similar foundation models in computer vision. These models could greatly simplify the use of images in any system by producing all-purpose visual features, i.e., features that work across image distributions and tasks without finetuning. This work shows that existing pretraining methods, especially self-supervised methods, can produce such features if trained on enough curated data from diverse sources. We revisit existing approaches and combine different techniques to scale our pretraining in terms of data and model size. Most of the technical contributions aim at accelerating and stabilizing the training at scale. In terms of data, we propose an automatic pipeline to build a dedicated, diverse, and curated image dataset instead of uncurated data, as typically done in the self-supervised literature. In terms of models, we train a ViT model (Dosovitskiy et al., 2020) with 1B parameters and distill it into a series of smaller models that surpass the best available all-purpose features, OpenCLIP (Ilharco et al., 2021) on most of the benchmarks at image and pixel levels.

Mahmoud Assran, Mathilde Caron, Ishan Misra et al. · meta-ai, mila

We propose Masked Siamese Networks (MSN), a self-supervised learning framework for learning image representations. Our approach matches the representation of an image view containing randomly masked patches to the representation of the original unmasked image. This self-supervised pre-training strategy is particularly scalable when applied to Vision Transformers since only the unmasked patches are processed by the network. As a result, MSNs improve the scalability of joint-embedding architectures, while producing representations of a high semantic level that perform competitively on low-shot image classification. For instance, on ImageNet-1K, with only 5,000 annotated images, our base MSN model achieves 72.4% top-1 accuracy, and with 1% of ImageNet-1K labels, we achieve 75.7% top-1 accuracy, setting a new state-of-the-art for self-supervised learning on this benchmark. Our code is publicly available.

Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero et al. · meta-ai

Deep learning vision systems are widely deployed across applications where reliability is critical. However, even today's best models can fail to recognize an object when its pose, lighting, or background varies. While existing benchmarks surface examples challenging for models, they do not explain why such mistakes arise. To address this need, we introduce ImageNet-X, a set of sixteen human annotations of factors such as pose, background, or lighting the entire ImageNet-1k validation set as well as a random subset of 12k training images. Equipped with ImageNet-X, we investigate 2,200 current recognition models and study the types of mistakes as a function of model's (1) architecture, e.g. transformer vs. convolutional, (2) learning paradigm, e.g. supervised vs. self-supervised, and (3) training procedures, e.g., data augmentation. Regardless of these choices, we find models have consistent failure modes across ImageNet-X categories. We also find that while data augmentation can improve robustness to certain factors, they induce spill-over effects to other factors. For example, strong random cropping hurts robustness on smaller objects. Together, these insights suggest to advance the robustness of modern vision models, future research should focus on collecting additional data and understanding data augmentation schemes. Along with these insights, we release a toolkit based on ImageNet-X to spur further study into the mistakes image recognition systems make.

Mahmoud Assran, Quentin Duval, Ishan Misra et al.

This paper demonstrates an approach for learning highly semantic image representations without relying on hand-crafted data-augmentations. We introduce the Image-based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. A core design choice to guide I-JEPA towards producing semantic representations is the masking strategy; specifically, it is crucial to (a) sample target blocks with sufficiently large scale (semantic), and to (b) use a sufficiently informative (spatially distributed) context block. Empirically, when combined with Vision Transformers, we find I-JEPA to be highly scalable. For instance, we train a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours to achieve strong downstream performance across a wide range of tasks, from linear classification to object counting and depth prediction.

Mohammad Pezeshki, Diane Bouchacourt, Mark Ibrahim et al.

Environment annotations are essential for the success of many out-of-distribution (OOD) generalization methods. Unfortunately, these are costly to obtain and often limited by human annotators' biases. To achieve robust generalization, it is essential to develop algorithms for automatic environment discovery within datasets. Current proposals, which divide examples based on their training error, suffer from one fundamental problem. These methods introduce hyper-parameters and early-stopping criteria, which require a validation set with human-annotated environments, the very information subject to discovery. In this paper, we propose Cross-Risk-Minimization (XRM) to address this issue. XRM trains twin networks, each learning from one random half of the training data, while imitating confident held-out mistakes made by its sibling. XRM provides a recipe for hyper-parameter tuning, does not require early-stopping, and can discover environments for all training and validation data. Algorithms built on top of XRM environments achieve oracle worst-group-accuracy, addressing a long-standing challenge in OOD generalization. Code available at \url{https://github.com/facebookresearch/XRM}.

Lorenzo Mur-Labadia, Matthew Muckley, Amir Bar et al. · meta-ai

We present V-JEPA 2.1, a family of self-supervised models that learn dense, high-quality visual representations for both images and videos while retaining strong global scene understanding. The approach combines four key components. First, a dense predictive loss uses a masking-based objective in which both visible and masked tokens contribute to the training signal, encouraging explicit spatial and temporal grounding. Second, deep self-supervision applies the self-supervised objective hierarchically across multiple intermediate encoder layers to improve representation quality. Third, multi-modal tokenizers enable unified training across images and videos. Finally, the model benefits from effective scaling in both model capacity and training data. Together, these design choices produce representations that are spatially structured, semantically coherent, and temporally consistent. Empirically, V-JEPA 2.1 achieves state-of-the-art performance on several challenging benchmarks, including 7.71 mAP on Ego4D for short-term object-interaction anticipation and 40.8 Recall@5 on EPIC-KITCHENS for high-level action anticipation, as well as a 20-point improvement in real-robot grasping success rate over V-JEPA-2 AC. The model also demonstrates strong performance in robotic navigation (5.687 ATE on TartanDrive), depth estimation (0.307 RMSE on NYUv2 with a linear probe), and global recognition (77.7 on Something-Something-V2). These results show that V-JEPA 2.1 significantly advances the state of the art in dense visual understanding and world modeling.

Mahmoud Assran, Randall Balestriero, Quentin Duval et al.

A successful paradigm in representation learning is to perform self-supervised pretraining using tasks based on mini-batch statistics (e.g., SimCLR, VICReg, SwAV, MSN). We show that in the formulation of all these methods is an overlooked prior to learn features that enable uniform clustering of the data. While this prior has led to remarkably semantic representations when pretraining on class-balanced data, such as ImageNet, we demonstrate that it can hamper performance when pretraining on class-imbalanced data. By moving away from conventional uniformity priors and instead preferring power-law distributed feature clusters, we show that one can improve the quality of the learned representations on real-world class-imbalanced datasets. To demonstrate this, we develop an extension of the Masked Siamese Networks (MSN) method to support the use of arbitrary features priors.

Nasim Rahaman, Martin Weiss, Francesco Locatello et al.

Recent work has seen the development of general purpose neural architectures that can be trained to perform tasks across diverse data modalities. General purpose models typically make few assumptions about the underlying data-structure and are known to perform well in the large-data regime. At the same time, there has been growing interest in modular neural architectures that represent the data using sparsely interacting modules. These models can be more robust out-of-distribution, computationally efficient, and capable of sample-efficient adaptation to new data. However, they tend to make domain-specific assumptions about the data, and present challenges in how module behavior (i.e., parameterization) and connectivity (i.e., their layout) can be jointly learned. In this work, we introduce a general purpose, yet modular neural architecture called Neural Attentive Circuits (NACs) that jointly learns the parameterization and a sparse connectivity of neural modules without using domain knowledge. NACs are best understood as the combination of two systems that are jointly trained end-to-end: one that determines the module configuration and the other that executes it on an input. We demonstrate qualitatively that NACs learn diverse and meaningful module configurations on the NLVR2 dataset without additional supervision. Quantitatively, we show that by incorporating modularity in this way, NACs improve upon a strong non-modular baseline in terms of low-shot adaptation on CIFAR and CUBs dataset by about 10%, and OOD robustness on Tiny ImageNet-R by about 2.5%. Further, we find that NACs can achieve an 8x speedup at inference time while losing less than 3% performance. Finally, we find NACs to yield competitive results on diverse data modalities spanning point-cloud classification, symbolic processing and text-classification from ASCII bytes, thereby confirming its general purpose nature.

Amir Bar, Florian Bordes, Assaf Shocher et al.

Masked Image Modeling (MIM) is a promising self-supervised learning approach that enables learning from unlabeled images. Despite its recent success, learning good representations through MIM remains challenging because it requires predicting the right semantic content in accurate locations. For example, given an incomplete picture of a dog, we can guess that there is a tail, but we cannot determine its exact location. In this work, we propose to incorporate location uncertainty into MIM by using stochastic positional embeddings (StoP). Specifically, we condition the model on stochastic masked token positions drawn from a Gaussian distribution. StoP reduces overfitting to location features and guides the model toward learning features that are more robust to location uncertainties. Quantitatively, StoP improves downstream MIM performance on a variety of downstream tasks, including $+1.7\%$ on ImageNet linear probing using ViT-B, and $+2.5\%$ for ViT-H using $1\%$ of the data.

Quentin Duval, Ishan Misra, Nicolas Ballas

Self-supervised methods in vision have been mostly focused on large architectures as they seem to suffer from a significant performance drop for smaller architectures. In this paper, we propose a simple self-supervised distillation technique that can train high performance low-compute neural networks. Our main insight is that existing joint-embedding based SSL methods can be repurposed for knowledge distillation from a large self-supervised teacher to a small student model. Thus, we call our method Replace one Branch (RoB) as it simply replaces one branch of the joint-embedding training with a large teacher model. RoB is widely applicable to a number of architectures such as small ResNets, MobileNets and ViT, and pretrained models such as DINO, SwAV or iBOT. When pretraining on the ImageNet dataset, RoB yields models that compete with supervised knowledge distillation. When applied to MSN, RoB produces students with strong semi-supervised capabilities. Finally, our best ViT-Tiny models improve over prior SSL state-of-the-art on ImageNet by $2.3\%$ and are on par or better than a supervised distilled DeiT on five downstream transfer tasks (iNaturalist, CIFAR, Clevr/Count, Clevr/Dist and Places). We hope RoB enables practical self-supervision at smaller scale.

Florian Bordes, Samuel Lavoie, Randall Balestriero et al.

Self-Supervised Learning (SSL) models rely on a pretext task to learn representations. Because this pretext task differs from the downstream tasks used to evaluate the performance of these models, there is an inherent misalignment or pretraining bias. A commonly used trick in SSL, shown to make deep networks more robust to such bias, is the addition of a small projector (usually a 2 or 3 layer multi-layer perceptron) on top of a backbone network during training. In contrast to previous work that studied the impact of the projector architecture, we here focus on a simpler, yet overlooked lever to control the information in the backbone representation. We show that merely changing its dimensionality -- by changing only the size of the backbone's very last block -- is a remarkably effective technique to mitigate the pretraining bias. It significantly improves downstream transfer performance for both Self-Supervised and Supervised pretrained models.

Lluis Castrejon, Nicolas Ballas, Aaron Courville

Videos can be created by first outlining a global view of the scene and then adding local details. Inspired by this idea we propose a cascaded model for video generation which follows a coarse to fine approach. First our model generates a low resolution video, establishing the global scene structure, which is then refined by subsequent cascade levels operating at larger resolutions. We train each cascade level sequentially on partial views of the videos, which reduces the computational complexity of our model and makes it scalable to high-resolution videos with many frames. We empirically validate our approach on UCF101 and Kinetics-600, for which our model is competitive with the state-of-the-art. We further demonstrate the scaling capabilities of our model and train a three-level model on the BDD100K dataset which generates 256x256 pixels videos with 48 frames.

Shengbang Tong, David Fan, John Nguyen et al.

The visual world offers a critical axis for advancing foundation models beyond language. Despite growing interest in this direction, the design space for native multimodal models remains opaque. We provide empirical clarity through controlled, from-scratch pretraining experiments, isolating the factors that govern multimodal pretraining without interference from language pretraining. We adopt the Transfusion framework, using next-token prediction for language and diffusion for vision, to train on diverse data including text, video, image-text pairs, and even action-conditioned video. Our experiments yield four key insights: (i) Representation Autoencoder (RAE) provides an optimal unified visual representation by excelling at both visual understanding and generation; (ii) visual and language data are complementary and yield synergy for downstream capabilities; (iii) unified multimodal pretraining leads naturally to world modeling, with capabilities emerging from general training; and (iv) Mixture-of-Experts (MoE) enables efficient and effective multimodal scaling while naturally inducing modality specialization. Through IsoFLOP analysis, we compute scaling laws for both modalities and uncover a scaling asymmetry: vision is significantly more data-hungry than language. We demonstrate that the MoE architecture harmonizes this scaling asymmetry by providing the high model capacity required by language while accommodating the data-intensive nature of vision, paving the way for truly unified multimodal models.

Jianhao Yuan, Xiaofeng Zhang, Felix Friedrich et al.

State-of-the-art video generative models produce promising visual content yet often violate basic physics principles, limiting their utility. While some attribute this deficiency to insufficient physics understanding from pre-training, we find that the shortfall in physics plausibility also stems from suboptimal inference strategies. We therefore introduce WMReward and treat improving physics plausibility of video generation as an inference-time alignment problem. In particular, we leverage the strong physics prior of a latent world model (here, VJEPA-2) as a reward to search and steer multiple candidate denoising trajectories, enabling scaling test-time compute for better generation performance. Empirically, our approach substantially improves physics plausibility across image-conditioned, multiframe-conditioned, and text-conditioned generation settings, with validation from human preference study. Notably, in the ICCV 2025 Perception Test PhysicsIQ Challenge, we achieve a final score of 62.64%, winning first place and outperforming the previous state of the art by 7.42%. Our work demonstrates the viability of using latent world models to improve physics plausibility of video generation, beyond this specific instantiation or parameterization.

Siddharth Verma, Yuchen Lu, Rui Hou et al.

Masked Language Modeling (MLM) has proven to be an essential component of Vision-Language (VL) pretraining. To implement MLM, the researcher must make two design choices: the masking strategy, which determines which tokens to mask, and the masking rate, which determines how many tokens to mask. Previous work has focused primarily on the masking strategy while setting the masking rate at a default of 15\%. In this paper, we show that increasing this masking rate improves downstream performance while simultaneously reducing performance gap among different masking strategies, rendering the uniform masking strategy competitive to other more complex ones. Surprisingly, we also discover that increasing the masking rate leads to gains in Image-Text Matching (ITM) tasks, suggesting that the role of MLM goes beyond language modeling in VL pretraining.

Quentin Garrido, Tushar Nagarajan, Basile Terver et al.

Agents capable of reasoning and planning in the real world require the ability of predicting the consequences of their actions. While world models possess this capability, they most often require action labels, that can be complex to obtain at scale. This motivates the learning of latent action models, that can learn an action space from videos alone. Our work addresses the problem of learning latent actions world models on in-the-wild videos, expanding the scope of existing works that focus on simple robotics simulations, video games, or manipulation data. While this allows us to capture richer actions, it also introduces challenges stemming from the video diversity, such as environmental noise, or the lack of a common embodiment across videos. To address some of the challenges, we discuss properties that actions should follow as well as relevant architectural choices and evaluations. We find that continuous, but constrained, latent actions are able to capture the complexity of actions from in-the-wild videos, something that the common vector quantization does not. We for example find that changes in the environment coming from agents, such as humans entering the room, can be transferred across videos. This highlights the capability of learning actions that are specific to in-the-wild videos. In the absence of a common embodiment across videos, we are mainly able to learn latent actions that become localized in space, relative to the camera. Nonetheless, we are able to train a controller that maps known actions to latent ones, allowing us to use latent actions as a universal interface and solve planning tasks with our world model with similar performance as action-conditioned baselines. Our analyses and experiments provide a step towards scaling latent action models to the real world.

Benno Krojer, Mojtaba Komeili, Candace Ross et al.

Existing benchmarks for assessing the spatio-temporal understanding and reasoning abilities of video language models are susceptible to score inflation due to the presence of shortcut solutions based on superficial visual or textual cues. This paper mitigates the challenges in accurately assessing model performance by introducing the Minimal Video Pairs (MVP) benchmark, a simple shortcut-aware video QA benchmark for assessing the physical understanding of video language models. The benchmark is comprised of 55K high-quality multiple-choice video QA examples focusing on physical world understanding. Examples are curated from nine video data sources, spanning first-person egocentric and exocentric videos, robotic interaction data, and cognitive science intuitive physics benchmarks. To mitigate shortcut solutions that rely on superficial visual or textual cues and biases, each sample in MVP has a minimal-change pair -- a visually similar video accompanied by an identical question but an opposing answer. To answer a question correctly, a model must provide correct answers for both examples in the minimal-change pair; as such, models that solely rely on visual or textual biases would achieve below random performance. Human performance on MVP is 92.9\%, while the best open-source state-of-the-art video-language model achieves 40.2\% compared to random performance at 25\%.

Adrien Bardes, Quentin Garrido, Jean Ponce et al.

This paper explores feature prediction as a stand-alone objective for unsupervised learning from video and introduces V-JEPA, a collection of vision models trained solely using a feature prediction objective, without the use of pretrained image encoders, text, negative examples, reconstruction, or other sources of supervision. The models are trained on 2 million videos collected from public datasets and are evaluated on downstream image and video tasks. Our results show that learning by predicting video features leads to versatile visual representations that perform well on both motion and appearance-based tasks, without adaption of the model's parameters; e.g., using a frozen backbone. Our largest model, a ViT-H/16 trained only on videos, obtains 81.9% on Kinetics-400, 72.2% on Something-Something-v2, and 77.9% on ImageNet1K.

Mahmoud Assran, Nicolas Ballas, Lluis Castrejon et al.

We investigate a strategy for improving the efficiency of contrastive learning of visual representations by leveraging a small amount of supervised information during pre-training. We propose a semi-supervised loss, SuNCEt, based on noise-contrastive estimation and neighbourhood component analysis, that aims to distinguish examples of different classes in addition to the self-supervised instance-wise pretext tasks. On ImageNet, we find that SuNCEt can be used to match the semi-supervised learning accuracy of previous contrastive approaches while using less than half the amount of pre-training and compute. Our main insight is that leveraging even a small amount of labeled data during pre-training, and not only during fine-tuning, provides an important signal that can significantly accelerate contrastive learning of visual representations. Our code is available online at github.com/facebookresearch/suncet.

Mido Assran, Adrien Bardes, David Fan et al. · meta-ai

A major challenge for modern AI is to learn to understand the world and learn to act largely by observation. This paper explores a self-supervised approach that combines internet-scale video data with a small amount of interaction data (robot trajectories), to develop models capable of understanding, predicting, and planning in the physical world. We first pre-train an action-free joint-embedding-predictive architecture, V-JEPA 2, on a video and image dataset comprising over 1 million hours of internet video. V-JEPA 2 achieves strong performance on motion understanding (77.3 top-1 accuracy on Something-Something v2) and state-of-the-art performance on human action anticipation (39.7 recall-at-5 on Epic-Kitchens-100) surpassing previous task-specific models. Additionally, after aligning V-JEPA 2 with a large language model, we demonstrate state-of-the-art performance on multiple video question-answering tasks at the 8 billion parameter scale (e.g., 84.0 on PerceptionTest, 76.9 on TempCompass). Finally, we show how self-supervised learning can be applied to robotic planning tasks by post-training a latent action-conditioned world model, V-JEPA 2-AC, using less than 62 hours of unlabeled robot videos from the Droid dataset. We deploy V-JEPA 2-AC zero-shot on Franka arms in two different labs and enable picking and placing of objects using planning with image goals. Notably, this is achieved without collecting any data from the robots in these environments, and without any task-specific training or reward. This work demonstrates how self-supervised learning from web-scale data and a small amount of robot interaction data can yield a world model capable of planning in the physical world.

Quentin Garrido, Mahmoud Assran, Nicolas Ballas et al.

Joint-Embedding Predictive Architecture (JEPA) has emerged as a promising self-supervised approach that learns by leveraging a world model. While previously limited to predicting missing parts of an input, we explore how to generalize the JEPA prediction task to a broader set of corruptions. We introduce Image World Models, an approach that goes beyond masked image modeling and learns to predict the effect of global photometric transformations in latent space. We study the recipe of learning performant IWMs and show that it relies on three key aspects: conditioning, prediction difficulty, and capacity. Additionally, we show that the predictive world model learned by IWM can be adapted through finetuning to solve diverse tasks; a fine-tuned IWM world model matches or surpasses the performance of previous self-supervised methods. Finally, we show that learning with an IWM allows one to control the abstraction level of the learned representations, learning invariant representations such as contrastive methods, or equivariant representations such as masked image modelling.

Shraman Pramanick, Guangxing Han, Rui Hou et al.

The ability of large language models (LLMs) to process visual inputs has given rise to general-purpose vision systems, unifying various vision-language (VL) tasks by instruction tuning. However, due to the enormous diversity in input-output formats in the vision domain, existing general-purpose models fail to successfully integrate segmentation and multi-image inputs with coarse-level tasks into a single framework. In this work, we introduce VistaLLM, a powerful visual system that addresses coarse- and fine-grained VL tasks over single and multiple input images using a unified framework. VistaLLM utilizes an instruction-guided image tokenizer that filters global embeddings using task descriptions to extract compressed and refined features from numerous images. Moreover, VistaLLM employs a gradient-aware adaptive sampling technique to represent binary segmentation masks as sequences, significantly improving over previously used uniform sampling. To bolster the desired capability of VistaLLM, we curate CoinIt, a comprehensive coarse-to-fine instruction tuning dataset with 6.8M samples. We also address the lack of multi-image grounding datasets by introducing a novel task, AttCoSeg (Attribute-level Co-Segmentation), which boosts the model's reasoning and grounding capability over multiple input images. Extensive experiments on a wide range of V- and VL tasks demonstrate the effectiveness of VistaLLM by achieving consistent state-of-the-art performance over strong baselines across all downstream tasks. Our project page can be found at https://shramanpramanick.github.io/VistaLLM/.

Samuel Lavoie, Polina Kirichenko, Mark Ibrahim et al.

There are a thousand ways to caption an image. Contrastive Language Pretraining (CLIP) on the other hand, works by mapping an image and its caption to a single vector -- limiting how well CLIP-like models can represent the diverse ways to describe an image. In this work, we introduce Llip, Latent Language Image Pretraining, which models the diversity of captions that could match an image. Llip's vision encoder outputs a set of visual features that are mixed into a final representation by conditioning on information derived from the text. We show that Llip outperforms non-contextualized baselines like CLIP and SigLIP on a variety of tasks even with large-scale encoders. Llip improves zero-shot classification by an average of 2.9% zero-shot classification benchmarks with a ViT-G/14 encoder. Specifically, Llip attains a zero-shot top-1 accuracy of 83.5% on ImageNet outperforming a similarly sized CLIP by 1.4%. We also demonstrate improvement on zero-shot retrieval on MS-COCO by 6.0%. We provide a comprehensive analysis of the components introduced by the method and demonstrate that Llip leads to richer visual representations.

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.

Quentin Garrido, Nicolas Ballas, Mahmoud Assran et al.

We investigate the emergence of intuitive physics understanding in general-purpose deep neural network models trained to predict masked regions in natural videos. Leveraging the violation-of-expectation framework, we find that video prediction models trained to predict outcomes in a learned representation space demonstrate an understanding of various intuitive physics properties, such as object permanence and shape consistency. In contrast, video prediction in pixel space and multimodal large language models, which reason through text, achieve performance closer to chance. Our comparisons of these architectures reveal that jointly learning an abstract representation space while predicting missing parts of sensory input, akin to predictive coding, is sufficient to acquire an understanding of intuitive physics, and that even models trained on one week of unique video achieve above chance performance. This challenges the idea that core knowledge -- a set of innate systems to help understand the world -- needs to be hardwired to develop an understanding of intuitive physics.

Sergio Arnaud, Paul McVay, Ada Martin et al. · mit

We present LOCATE 3D, a model for localizing objects in 3D scenes from referring expressions like "the small coffee table between the sofa and the lamp." LOCATE 3D sets a new state-of-the-art on standard referential grounding benchmarks and showcases robust generalization capabilities. Notably, LOCATE 3D operates directly on sensor observation streams (posed RGB-D frames), enabling real-world deployment on robots and AR devices. Key to our approach is 3D-JEPA, a novel self-supervised learning (SSL) algorithm applicable to sensor point clouds. It takes as input a 3D pointcloud featurized using 2D foundation models (CLIP, DINO). Subsequently, masked prediction in latent space is employed as a pretext task to aid the self-supervised learning of contextualized pointcloud features. Once trained, the 3D-JEPA encoder is finetuned alongside a language-conditioned decoder to jointly predict 3D masks and bounding boxes. Additionally, we introduce LOCATE 3D DATASET, a new dataset for 3D referential grounding, spanning multiple capture setups with over 130K annotations. This enables a systematic study of generalization capabilities as well as a stronger model.

Wancong Zhang, Basile Terver, Artem Zholus et al.

Model predictive control (MPC) with learned world models has emerged as a promising paradigm for embodied control, particularly for its ability to generalize zero-shot when deployed in new environments. However, learned world models often struggle with long-horizon control due to the accumulation of prediction errors and the exponentially growing search space. In this work, we address these challenges by learning latent world models at multiple temporal scales and performing hierarchical planning across these scales, enabling long-horizon reasoning while substantially reducing inference-time planning complexity. Our approach serves as a modular planning abstraction that applies across diverse latent world-model architectures and domains. We demonstrate that this hierarchical approach enables zero-shot control on real-world non-greedy robotic tasks, achieving a 70% success rate on pick-&-place using only a final goal specification, compared to 0% for a single-level world model. In addition, across physics-based simulated environments including push manipulation and maze navigation, hierarchical planning achieves higher success while requiring up to 4x less planning-time compute.

Aniket Didolkar, Nicolas Ballas, Sanjeev Arora et al.

Large language models (LLMs) now solve multi-step problems by emitting extended chains of thought. During the process, they often re-derive the same intermediate steps across problems, inflating token usage and latency. This saturation of the context window leaves less capacity for exploration. We study a simple mechanism that converts recurring reasoning fragments into concise, reusable "behaviors" (name + instruction) via the model's own metacognitive analysis of prior traces. These behaviors are stored in a "behavior handbook" which supplies them to the model in-context at inference or distills them into parameters via supervised fine-tuning. This approach achieves improved test-time reasoning across three different settings - 1) Behavior-conditioned inference: Providing the LLM relevant behaviors in-context during reasoning reduces number of reasoning tokens by up to 46% while matching or improving baseline accuracy; 2) Behavior-guided self-improvement: Without any parameter updates, the model improves its own future reasoning by leveraging behaviors from its own past problem solving attempts. This yields up to 10% higher accuracy than a naive critique-and-revise baseline; and 3) Behavior-conditioned SFT: SFT on behavior-conditioned reasoning traces is more effective at converting non-reasoning models into reasoning models as compared to vanilla SFT. Together, these results indicate that turning slow derivations into fast procedural hints enables LLMs to remember how to reason, not just what to conclude.

Randall Balestriero, Nicolas Ballas, Mike Rabbat et al.

Joint Embedding Predictive Architectures (JEPAs) learn representations able to solve numerous downstream tasks out-of-the-box. JEPAs combine two objectives: (i) a latent-space prediction term, i.e., the representation of a slightly perturbed sample must be predictable from the original sample's representation, and (ii) an anti-collapse term, i.e., not all samples should have the same representation. While (ii) is often considered as an obvious remedy to representation collapse, we uncover that JEPAs' anti-collapse term does much more--it provably estimates the data density. In short, any successfully trained JEPA can be used to get sample probabilities, e.g., for data curation, outlier detection, or simply for density estimation. Our theoretical finding is agnostic of the dataset and architecture used--in any case one can compute the learned probabilities of sample $x$ efficiently and in closed-form using the model's Jacobian matrix at $x$. Our findings are empirically validated across datasets (synthetic, controlled, and Imagenet) and across different Self Supervised Learning methods falling under the JEPA family (I-JEPA and DINOv2) and on multimodal models, such as MetaCLIP. We denote the method extracting the JEPA learned density as {\bf JEPA-SCORE}.

Jianhao Yuan, Xiaofeng Zhang, Felix Friedrich et al.

This is a short technical report describing the winning entry of the PhysicsIQ Challenge, presented at the Perception Test Workshop at ICCV 2025. State-of-the-art video generative models exhibit severely limited physical understanding, and often produce implausible videos. The Physics IQ benchmark has shown that visual realism does not imply physics understanding. Yet, intuitive physics understanding has shown to emerge from SSL pretraining on natural videos. In this report, we investigate whether we can leverage SSL-based video world models to improve the physics plausibility of video generative models. In particular, we build ontop of the state-of-the-art video generative model MAGI-1 and couple it with the recently introduced Video Joint Embedding Predictive Architecture 2 (VJEPA-2) to guide the generation process. We show that by leveraging VJEPA-2 as reward signal, we can improve the physics plausibility of state-of-the-art video generative models by ~6%.

Nimit S. Sohoni, Maziar Sanjabi, Nicolas Ballas et al.

While neural networks have shown remarkable success on classification tasks in terms of average-case performance, they often fail to perform well on certain groups of the data. Such group information may be expensive to obtain; thus, recent works in robustness and fairness have proposed ways to improve worst-group performance even when group labels are unavailable for the training data. However, these methods generally underperform methods that utilize group information at training time. In this work, we assume access to a small number of group labels alongside a larger dataset without group labels. We propose BARACK, a simple two-step framework to utilize this partial group information to improve worst-group performance: train a model to predict the missing group labels for the training data, and then use these predicted group labels in a robust optimization objective. Theoretically, we provide generalization bounds for our approach in terms of the worst-group performance, which scale with respect to both the total number of training points and the number of training points with group labels. Empirically, our method outperforms the baselines that do not use group information, even when only 1-33% of points have group labels. We provide ablation studies to support the robustness and extensibility of our framework.

Jose Javier Gonzalez Ortiz, Jonathan Frankle, Mike Rabbat et al.

As datasets and models become increasingly large, distributed training has become a necessary component to allow deep neural networks to train in reasonable amounts of time. However, distributed training can have substantial communication overhead that hinders its scalability. One strategy for reducing this overhead is to perform multiple unsynchronized SGD steps independently on each worker between synchronization steps, a technique known as local SGD. We conduct a comprehensive empirical study of local SGD and related methods on a large-scale image classification task. We find that performing local SGD comes at a price: lower communication costs (and thereby faster training) are accompanied by lower accuracy. This finding is in contrast from the smaller-scale experiments in prior work, suggesting that local SGD encounters challenges at scale. We further show that incorporating the slow momentum framework of Wang et al. (2020) consistently improves accuracy without requiring additional communication, hinting at future directions for potentially escaping this trade-off.

Lluis Castrejon, Nicolas Ballas, Aaron Courville

Videos can often be created by first outlining a global description of the scene and then adding local details. Inspired by this we propose a hierarchical model for video generation which follows a coarse to fine approach. First our model generates a low resolution video, establishing the global scene structure, that is then refined by subsequent levels in the hierarchy. We train each level in our hierarchy sequentially on partial views of the videos. This reduces the computational complexity of our generative model, which scales to high-resolution videos beyond a few frames. We validate our approach on Kinetics-600 and BDD100K, for which we train a three level model capable of generating 256x256 videos with 48 frames.

Mahmoud Assran, Mathilde Caron, Ishan Misra et al.

This paper proposes a novel method of learning by predicting view assignments with support samples (PAWS). The method trains a model to minimize a consistency loss, which ensures that different views of the same unlabeled instance are assigned similar pseudo-labels. The pseudo-labels are generated non-parametrically, by comparing the representations of the image views to those of a set of randomly sampled labeled images. The distance between the view representations and labeled representations is used to provide a weighting over class labels, which we interpret as a soft pseudo-label. By non-parametrically incorporating labeled samples in this way, PAWS extends the distance-metric loss used in self-supervised methods such as BYOL and SwAV to the semi-supervised setting. Despite the simplicity of the approach, PAWS outperforms other semi-supervised methods across architectures, setting a new state-of-the-art for a ResNet-50 on ImageNet trained with either 10% or 1% of the labels, reaching 75.5% and 66.5% top-1 respectively. PAWS requires 4x to 12x less training than the previous best methods.

Shagun Sodhani, Olivier Delalleau, Mahmoud Assran et al.

Codistillation has been proposed as a mechanism to share knowledge among concurrently trained models by encouraging them to represent the same function through an auxiliary loss. This contrasts with the more commonly used fully-synchronous data-parallel stochastic gradient descent methods, where different model replicas average their gradients (or parameters) at every iteration and thus maintain identical parameters. We investigate codistillation in a distributed training setup, complementing previous work which focused on extremely large batch sizes. Surprisingly, we find that even at moderate batch sizes, models trained with codistillation can perform as well as models trained with synchronous data-parallel methods, despite using a much weaker synchronization mechanism. These findings hold across a range of batch sizes and learning rate schedules, as well as different kinds of models and datasets. Obtaining this level of accuracy, however, requires properly accounting for the regularization effect of codistillation, which we highlight through several empirical observations. Overall, this work contributes to a better understanding of codistillation and how to best take advantage of it in a distributed computing environment.

César Laurent, Camille Ballas, Thomas George et al.

By removing parameters from deep neural networks, unstructured pruning methods aim at cutting down memory footprint and computational cost, while maintaining prediction accuracy. In order to tackle this otherwise intractable problem, many of these methods model the loss landscape using first or second order Taylor expansions to identify which parameters can be discarded. We revisit loss modelling for unstructured pruning: we show the importance of ensuring locality of the pruning steps. We systematically compare first and second order Taylor expansions and empirically show that both can reach similar levels of performance. Finally, we show that better preserving the original network function does not necessarily transfer to better performing networks after fine-tuning, suggesting that only considering the impact of pruning on the loss might not be a sufficient objective to design good pruning criteria.

Jianyu Wang, Vinayak Tantia, Nicolas Ballas et al.

Distributed optimization is essential for training large models on large datasets. Multiple approaches have been proposed to reduce the communication overhead in distributed training, such as synchronizing only after performing multiple local SGD steps, and decentralized methods (e.g., using gossip algorithms) to decouple communications among workers. Although these methods run faster than AllReduce-based methods, which use blocking communication before every update, the resulting models may be less accurate after the same number of updates. Inspired by the BMUF method of Chen & Huo (2016), we propose a slow momentum (SlowMo) framework, where workers periodically synchronize and perform a momentum update, after multiple iterations of a base optimization algorithm. Experiments on image classification and machine translation tasks demonstrate that SlowMo consistently yields improvements in optimization and generalization performance relative to the base optimizer, even when the additional overhead is amortized over many updates so that the SlowMo runtime is on par with that of the base optimizer. We provide theoretical convergence guarantees showing that SlowMo converges to a stationary point of smooth non-convex losses. Since BMUF can be expressed through the SlowMo framework, our results also correspond to the first theoretical convergence guarantees for BMUF.

Nick Pawlowski, Suvrat Bhooshan, Nicolas Ballas et al.

In some important computer vision domains, such as medical or hyperspectral imaging, we care about the classification of tiny objects in large images. However, most Convolutional Neural Networks (CNNs) for image classification were developed using biased datasets that contain large objects, in mostly central image positions. To assess whether classical CNN architectures work well for tiny object classification we build a comprehensive testbed containing two datasets: one derived from MNIST digits and one from histopathology images. This testbed allows controlled experiments to stress-test CNN architectures with a broad spectrum of signal-to-noise ratios. Our observations indicate that: (1) There exists a limit to signal-to-noise below which CNNs fail to generalize and that this limit is affected by dataset size - more data leading to better performances; however, the amount of training data required for the model to generalize scales rapidly with the inverse of the object-to-image ratio (2) in general, higher capacity models exhibit better generalization; (3) when knowing the approximate object sizes, adapting receptive field is beneficial; and (4) for very small signal-to-noise ratio the choice of global pooling operation affects optimization, whereas for relatively large signal-to-noise values, all tested global pooling operations exhibit similar performance.

Mahmoud Assran, Joshua Romoff, Nicolas Ballas et al.

Multi-simulator training has contributed to the recent success of Deep Reinforcement Learning by stabilizing learning and allowing for higher training throughputs. We propose Gossip-based Actor-Learner Architectures (GALA) where several actor-learners (such as A2C agents) are organized in a peer-to-peer communication topology, and exchange information through asynchronous gossip in order to take advantage of a large number of distributed simulators. We prove that GALA agents remain within an epsilon-ball of one-another during training when using loosely coupled asynchronous communication. By reducing the amount of synchronization between agents, GALA is more computationally efficient and scalable compared to A2C, its fully-synchronous counterpart. GALA also outperforms A2C, being more robust and sample efficient. We show that we can run several loosely coupled GALA agents in parallel on a single GPU and achieve significantly higher hardware utilization and frame-rates than vanilla A2C at comparable power draws.

Lluis Castrejon, Nicolas Ballas, Aaron Courville

Predicting future frames for a video sequence is a challenging generative modeling task. Promising approaches include probabilistic latent variable models such as the Variational Auto-Encoder. While VAEs can handle uncertainty and model multiple possible future outcomes, they have a tendency to produce blurry predictions. In this work we argue that this is a sign of underfitting. To address this issue, we propose to increase the expressiveness of the latent distributions and to use higher capacity likelihood models. Our approach relies on a hierarchy of latent variables, which defines a family of flexible prior and posterior distributions in order to better model the probability of future sequences. We validate our proposal through a series of ablation experiments and compare our approach to current state-of-the-art latent variable models. Our method performs favorably under several metrics in three different datasets.

Mahmoud Assran, Nicolas Loizou, Nicolas Ballas et al.

Distributed data-parallel algorithms aim to accelerate the training of deep neural networks by parallelizing the computation of large mini-batch gradient updates across multiple nodes. Approaches that synchronize nodes using exact distributed averaging (e.g., via AllReduce) are sensitive to stragglers and communication delays. The PushSum gossip algorithm is robust to these issues, but only performs approximate distributed averaging. This paper studies Stochastic Gradient Push (SGP), which combines PushSum with stochastic gradient updates. We prove that SGP converges to a stationary point of smooth, non-convex objectives at the same sub-linear rate as SGD, and that all nodes achieve consensus. We empirically validate the performance of SGP on image classification (ResNet-50, ImageNet) and machine translation (Transformer, WMT'16 En-De) workloads. Our code will be made publicly available.

Stanisław Jastrzębski, Zachary Kenton, Nicolas Ballas et al.

Stochastic Gradient Descent (SGD) based training of neural networks with a large learning rate or a small batch-size typically ends in well-generalizing, flat regions of the weight space, as indicated by small eigenvalues of the Hessian of the training loss. However, the curvature along the SGD trajectory is poorly understood. An empirical investigation shows that initially SGD visits increasingly sharp regions, reaching a maximum sharpness determined by both the learning rate and the batch-size of SGD. When studying the SGD dynamics in relation to the sharpest directions in this initial phase, we find that the SGD step is large compared to the curvature and commonly fails to minimize the loss along the sharpest directions. Furthermore, using a reduced learning rate along these directions can improve training speed while leading to both sharper and better generalizing solutions compared to vanilla SGD. In summary, our analysis of the dynamics of SGD in the subspace of the sharpest directions shows that they influence the regions that SGD steers to (where larger learning rate or smaller batch size result in wider regions visited), the overall training speed, and the generalization ability of the final model.

Amy Zhang, Nicolas Ballas, Joelle Pineau

The risks and perils of overfitting in machine learning are well known. However most of the treatment of this, including diagnostic tools and remedies, was developed for the supervised learning case. In this work, we aim to offer new perspectives on the characterization and prevention of overfitting in deep Reinforcement Learning (RL) methods, with a particular focus on continuous domains. We examine several aspects, such as how to define and diagnose overfitting in MDPs, and how to reduce risks by injecting sufficient training diversity. This work complements recent findings on the brittleness of deep RL methods and offers practical observations for RL researchers and practitioners.

Thomas George, César Laurent, Xavier Bouthillier et al.

Optimization algorithms that leverage gradient covariance information, such as variants of natural gradient descent (Amari, 1998), offer the prospect of yielding more effective descent directions. For models with many parameters, the covariance matrix they are based on becomes gigantic, making them inapplicable in their original form. This has motivated research into both simple diagonal approximations and more sophisticated factored approximations such as KFAC (Heskes, 2000; Martens & Grosse, 2015; Grosse & Martens, 2016). In the present work we draw inspiration from both to propose a novel approximation that is provably better than KFAC and amendable to cheap partial updates. It consists in tracking a diagonal variance, not in parameter coordinates, but in a Kronecker-factored eigenbasis, in which the diagonal approximation is likely to be more effective. Experiments show improvements over KFAC in optimization speed for several deep network architectures.

Stanisław Jastrzębski, Zachary Kenton, Devansh Arpit et al.

We investigate the dynamical and convergent properties of stochastic gradient descent (SGD) applied to Deep Neural Networks (DNNs). Characterizing the relation between learning rate, batch size and the properties of the final minima, such as width or generalization, remains an open question. In order to tackle this problem we investigate the previously proposed approximation of SGD by a stochastic differential equation (SDE). We theoretically argue that three factors - learning rate, batch size and gradient covariance - influence the minima found by SGD. In particular we find that the ratio of learning rate to batch size is a key determinant of SGD dynamics and of the width of the final minima, and that higher values of the ratio lead to wider minima and often better generalization. We confirm these findings experimentally. Further, we include experiments which show that learning rate schedules can be replaced with batch size schedules and that the ratio of learning rate to batch size is an important factor influencing the memorization process.

Stanisław Jastrzębski, Devansh Arpit, Nicolas Ballas et al.

Residual networks (Resnets) have become a prominent architecture in deep learning. However, a comprehensive understanding of Resnets is still a topic of ongoing research. A recent view argues that Resnets perform iterative refinement of features. We attempt to further expose properties of this aspect. To this end, we study Resnets both analytically and empirically. We formalize the notion of iterative refinement in Resnets by showing that residual connections naturally encourage features of residual blocks to move along the negative gradient of loss as we go from one block to the next. In addition, our empirical analysis suggests that Resnets are able to perform both representation learning and iterative refinement. In general, a Resnet block tends to concentrate representation learning behavior in the first few layers while higher layers perform iterative refinement of features. Finally we observe that sharing residual layers naively leads to representation explosion and counterintuitively, overfitting, and we show that simple existing strategies can help alleviating this problem.

Devansh Arpit, Stanisław Jastrzębski, Nicolas Ballas et al.

We examine the role of memorization in deep learning, drawing connections to capacity, generalization, and adversarial robustness. While deep networks are capable of memorizing noise data, our results suggest that they tend to prioritize learning simple patterns first. In our experiments, we expose qualitative differences in gradient-based optimization of deep neural networks (DNNs) on noise vs. real data. We also demonstrate that for appropriately tuned explicit regularization (e.g., dropout) we can degrade DNN training performance on noise datasets without compromising generalization on real data. Our analysis suggests that the notions of effective capacity which are dataset independent are unlikely to explain the generalization performance of deep networks when trained with gradient based methods because training data itself plays an important role in determining the degree of memorization.

Tegan Maharaj, Nicolas Ballas, Anna Rohrbach et al.

While deep convolutional neural networks frequently approach or exceed human-level performance at benchmark tasks involving static images, extending this success to moving images is not straightforward. Having models which can learn to understand video is of interest for many applications, including content recommendation, prediction, summarization, event/object detection and understanding human visual perception, but many domains lack sufficient data to explore and perfect video models. In order to address the need for a simple, quantitative benchmark for developing and understanding video, we present MovieFIB, a fill-in-the-blank question-answering dataset with over 300,000 examples, based on descriptive video annotations for the visually impaired. In addition to presenting statistics and a description of the dataset, we perform a detailed analysis of 5 different models' predictions, and compare these with human performance. We investigate the relative importance of language, static (2D) visual features, and moving (3D) visual features; the effects of increasing dataset size, the number of frames sampled; and of vocabulary size. We illustrate that: this task is not solvable by a language model alone; our model combining 2D and 3D visual information indeed provides the best result; all models perform significantly worse than human-level. We provide human evaluations for responses given by different models and find that accuracy on the MovieFIB evaluation corresponds well with human judgement. We suggest avenues for improving video models, and hope that the proposed dataset can be useful for measuring and encouraging progress in this very interesting field.

David Krueger, Tegan Maharaj, János Kramár et al.

We propose zoneout, a novel method for regularizing RNNs. At each timestep, zoneout stochastically forces some hidden units to maintain their previous values. Like dropout, zoneout uses random noise to train a pseudo-ensemble, improving generalization. But by preserving instead of dropping hidden units, gradient information and state information are more readily propagated through time, as in feedforward stochastic depth networks. We perform an empirical investigation of various RNN regularizers, and find that zoneout gives significant performance improvements across tasks. We achieve competitive results with relatively simple models in character- and word-level language modelling on the Penn Treebank and Text8 datasets, and combining with recurrent batch normalization yields state-of-the-art results on permuted sequential MNIST.

The Theano Development Team, Rami Al-Rfou, Guillaume Alain et al.

Theano is a Python library that allows to define, optimize, and evaluate mathematical expressions involving multi-dimensional arrays efficiently. Since its introduction, it has been one of the most used CPU and GPU mathematical compilers - especially in the machine learning community - and has shown steady performance improvements. Theano is being actively and continuously developed since 2008, multiple frameworks have been built on top of it and it has been used to produce many state-of-the-art machine learning models. The present article is structured as follows. Section I provides an overview of the Theano software and its community. Section II presents the principal features of Theano and how to use them, and compares them with other similar projects. Section III focuses on recently-introduced functionalities and improvements. Section IV compares the performance of Theano against Torch7 and TensorFlow on several machine learning models. Section V discusses current limitations of Theano and potential ways of improving it.