Naila Murray

FAIR CodeGen team, Jade Copet, Quentin Carbonneaux et al. · meta-ai

We release Code World Model (CWM), a 32-billion-parameter open-weights LLM, to advance research on code generation with world models. To improve code understanding beyond what can be learned from training on static code alone, we mid-train CWM on a large amount of observation-action trajectories from Python interpreter and agentic Docker environments, and perform extensive multi-task reasoning RL in verifiable coding, math, and multi-turn software engineering environments. With CWM, we provide a strong testbed for researchers to explore the opportunities world modeling affords for improving code generation with reasoning and planning in computational environments. We present first steps of how world models can benefit agentic coding, enable step-by-step simulation of Python code execution, and show early results of how reasoning can benefit from the latter. CWM is a dense, decoder-only LLM trained with a context size of up to 131k tokens. Independent of its world modeling capabilities, CWM offers strong performance on general coding and math tasks: it reaches pass@1 scores of 65.8% on SWE-bench Verified (with test-time scaling), 68.6% on LiveCodeBench, 96.6% on Math-500, and 76.0% on AIME 2024. To support further research on code world modeling, we release model checkpoints after mid-training, SFT, and RL.

Eric Hambro, Roberta Raileanu, Danielle Rothermel et al.

Recent breakthroughs in the development of agents to solve challenging sequential decision making problems such as Go, StarCraft, or DOTA, have relied on both simulated environments and large-scale datasets. However, progress on this research has been hindered by the scarcity of open-sourced datasets and the prohibitive computational cost to work with them. Here we present the NetHack Learning Dataset (NLD), a large and highly-scalable dataset of trajectories from the popular game of NetHack, which is both extremely challenging for current methods and very fast to run. NLD consists of three parts: 10 billion state transitions from 1.5 million human trajectories collected on the NAO public NetHack server from 2009 to 2020; 3 billion state-action-score transitions from 100,000 trajectories collected from the symbolic bot winner of the NetHack Challenge 2021; and, accompanying code for users to record, load and stream any collection of such trajectories in a highly compressed form. We evaluate a wide range of existing algorithms including online and offline RL, as well as learning from demonstrations, showing that significant research advances are needed to fully leverage large-scale datasets for challenging sequential decision making tasks.

Dahyun Kang, Piotr Koniusz, Minsu Cho et al.

We address the task of weakly-supervised few-shot image classification and segmentation, by leveraging a Vision Transformer (ViT) pretrained with self-supervision. Our proposed method takes token representations from the self-supervised ViT and leverages their correlations, via self-attention, to produce classification and segmentation predictions through separate task heads. Our model is able to effectively learn to perform classification and segmentation in the absence of pixel-level labels during training, using only image-level labels. To do this it uses attention maps, created from tokens generated by the self-supervised ViT backbone, as pixel-level pseudo-labels. We also explore a practical setup with ``mixed" supervision, where a small number of training images contains ground-truth pixel-level labels and the remaining images have only image-level labels. For this mixed setup, we propose to improve the pseudo-labels using a pseudo-label enhancer that was trained using the available ground-truth pixel-level labels. Experiments on Pascal-5i and COCO-20i demonstrate significant performance gains in a variety of supervision settings, and in particular when little-to-no pixel-level labels are available.

Shan Zhang, Naila Murray, Lei Wang et al.

In this paper, we tackle the challenging problem of Few-shot Object Detection. Existing FSOD pipelines (i) use average-pooled representations that result in information loss; and/or (ii) discard position information that can help detect object instances. Consequently, such pipelines are sensitive to large intra-class appearance and geometric variations between support and query images. To address these drawbacks, we propose a Time-rEversed diffusioN tEnsor Transformer (TENET), which i) forms high-order tensor representations that capture multi-way feature occurrences that are highly discriminative, and ii) uses a transformer that dynamically extracts correlations between the query image and the entire support set, instead of a single average-pooled support embedding. We also propose a Transformer Relation Head (TRH), equipped with higher-order representations, which encodes correlations between query regions and the entire support set, while being sensitive to the positional variability of object instances. Our model achieves state-of-the-art results on PASCAL VOC, FSOD, and COCO.

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.

M. Saquib Sarfraz, Naila Murray, Vivek Sharma et al.

Action segmentation refers to inferring boundaries of semantically consistent visual concepts in videos and is an important requirement for many video understanding tasks. For this and other video understanding tasks, supervised approaches have achieved encouraging performance but require a high volume of detailed frame-level annotations. We present a fully automatic and unsupervised approach for segmenting actions in a video that does not require any training. Our proposal is an effective temporally-weighted hierarchical clustering algorithm that can group semantically consistent frames of the video. Our main finding is that representing a video with a 1-nearest neighbor graph by taking into account the time progression is sufficient to form semantically and temporally consistent clusters of frames where each cluster may represent some action in the video. Additionally, we establish strong unsupervised baselines for action segmentation and show significant performance improvements over published unsupervised methods on five challenging action segmentation datasets. Our code is available at https://github.com/ssarfraz/FINCH-Clustering/tree/master/TW-FINCH

Pierre-Emmanuel Mazaré, Gergely Szilvasy, Maria Lomeli et al.

Self-attention in transformer models is an incremental associative memory that maps key vectors to value vectors. One way to speed up self-attention is to employ GPU-compatible vector search algorithms based on standard partitioning methods such as k-means. However, such partitioning methods yield poor results in this context because (1) the keys and queries follow different distributions, and (2) the RoPE positional encoding hinders the bucket assignment. This paper introduces Saap (Self-Attention with Asymmetric Partitions), which overcomes these problems. It is an asymmetrical indexing technique that employs distinct partitions for keys and queries, thereby approximating self-attention with a data-adaptive sparsity pattern. It works on pretrained language models and only requires to train (offline) a small query classifier. On a long context Llama 3.1-8b model, with sequences ranging from 100k to 500k tokens, Saap typically reduces by a factor of 20 the fraction of memory that needs to be looked-up, which translates to a time saving of 60\% when compared to FlashAttention-v2.

Zheng Zhao, Yeskendir Koishekenov, Xianjun Yang et al.

Current Chain-of-Thought (CoT) verification methods predict reasoning correctness based on outputs (black-box) or activations (gray-box), but offer limited insight into why a computation fails. We introduce a white-box method: Circuit-based Reasoning Verification (CRV). We hypothesize that attribution graphs of correct CoT steps, viewed as execution traces of the model's latent reasoning circuits, possess distinct structural fingerprints from those of incorrect steps. By training a classifier on structural features of these graphs, we show that these traces contain a powerful signal of reasoning errors. Our white-box approach yields novel scientific insights unattainable by other methods. (1) We demonstrate that structural signatures of error are highly predictive, establishing the viability of verifying reasoning directly via its computational graph. (2) We find these signatures to be highly domain-specific, revealing that failures in different reasoning tasks manifest as distinct computational patterns. (3) We provide evidence that these signatures are not merely correlational; by using our analysis to guide targeted interventions on individual transcoder features, we successfully correct the model's faulty reasoning. Our work shows that, by scrutinizing a model's computational process, we can move from simple error detection to a deeper, causal understanding of LLM reasoning.

Yohann Cabon, Naila Murray, Martin Humenberger

This paper introduces an updated version of the well-known Virtual KITTI dataset which consists of 5 sequence clones from the KITTI tracking benchmark. In addition, the dataset provides different variants of these sequences such as modified weather conditions (e.g. fog, rain) or modified camera configurations (e.g. rotated by 15 degrees). For each sequence, we provide multiple sets of images containing RGB, depth, class segmentation, instance segmentation, flow, and scene flow data. Camera parameters and poses as well as vehicle locations are available as well. In order to showcase some of the dataset's capabilities, we ran multiple relevant experiments using state-of-the-art algorithms from the field of autonomous driving. The dataset is available for download at https://europe.naverlabs.com/Research/Computer-Vision/Proxy-Virtual-Worlds.

César Roberto de Souza, Adrien Gaidon, Yohann Cabon et al.

Deep video action recognition models have been highly successful in recent years but require large quantities of manually annotated data, which are expensive and laborious to obtain. In this work, we investigate the generation of synthetic training data for video action recognition, as synthetic data have been successfully used to supervise models for a variety of other computer vision tasks. We propose an interpretable parametric generative model of human action videos that relies on procedural generation, physics models and other components of modern game engines. With this model we generate a diverse, realistic, and physically plausible dataset of human action videos, called PHAV for "Procedural Human Action Videos". PHAV contains a total of 39,982 videos, with more than 1,000 examples for each of 35 action categories. Our video generation approach is not limited to existing motion capture sequences: 14 of these 35 categories are procedurally defined synthetic actions. In addition, each video is represented with 6 different data modalities, including RGB, optical flow and pixel-level semantic labels. These modalities are generated almost simultaneously using the Multiple Render Targets feature of modern GPUs. In order to leverage PHAV, we introduce a deep multi-task (i.e. that considers action classes from multiple datasets) representation learning architecture that is able to simultaneously learn from synthetic and real video datasets, even when their action categories differ. Our experiments on the UCF-101 and HMDB-51 benchmarks suggest that combining our large set of synthetic videos with small real-world datasets can boost recognition performance. Our approach also significantly outperforms video representations produced by fine-tuning state-of-the-art unsupervised generative models of videos.

Saumya Jetley, Naila Murray, Eleonora Vig

Most saliency estimation methods aim to explicitly model low-level conspicuity cues such as edges or blobs and may additionally incorporate top-down cues using face or text detection. Data-driven methods for training saliency models using eye-fixation data are increasingly popular, particularly with the introduction of large-scale datasets and deep architectures. However, current methods in this latter paradigm use loss functions designed for classification or regression tasks whereas saliency estimation is evaluated on topographical maps. In this work, we introduce a new saliency map model which formulates a map as a generalized Bernoulli distribution. We then train a deep architecture to predict such maps using novel loss functions which pair the softmax activation function with measures designed to compute distances between probability distributions. We show in extensive experiments the effectiveness of such loss functions over standard ones on four public benchmark datasets, and demonstrate improved performance over state-of-the-art saliency methods.

Jon Almazan, Bojana Gajic, Naila Murray et al.

Training a deep architecture using a ranking loss has become standard for the person re-identification task. Increasingly, these deep architectures include additional components that leverage part detections, attribute predictions, pose estimators and other auxiliary information, in order to more effectively localize and align discriminative image regions. In this paper we adopt a different approach and carefully design each component of a simple deep architecture and, critically, the strategy for training it effectively for person re-identification. We extensively evaluate each design choice, leading to a list of good practices for person re-identification. By following these practices, our approach outperforms the state of the art, including more complex methods with auxiliary components, by large margins on four benchmark datasets. We also provide a qualitative analysis of our trained representation which indicates that, while compact, it is able to capture information from localized and discriminative regions, in a manner akin to an implicit attention mechanism.

Naila Murray, Albert Gordo

Image aesthetics has become an important criterion for visual content curation on social media sites and media content repositories. Previous work on aesthetic prediction models in the computer vision community has focused on aesthetic score prediction or binary image labeling. However, raw aesthetic annotations are in the form of score histograms and provide richer and more precise information than binary labels or mean scores. Consequently, in this work we focus on the rarely-studied problem of predicting aesthetic score distributions and propose a novel architecture and training procedure for our model. Our model achieves state-of-the-art results on the standard AVA large-scale benchmark dataset for three tasks: (i) aesthetic quality classification; (ii) aesthetic score regression; and (iii) aesthetic score distribution prediction, all while using one model trained only for the distribution prediction task. We also introduce a method to modify an image such that its predicted aesthetics changes, and use this modification to gain insight into our model.

Naila Murray, Hervé Jégou, Florent Perronnin et al.

We consider the design of an image representation that embeds and aggregates a set of local descriptors into a single vector. Popular representations of this kind include the bag-of-visual-words, the Fisher vector and the VLAD. When two such image representations are compared with the dot-product, the image-to-image similarity can be interpreted as a match kernel. In match kernels, one has to deal with interference, i.e. with the fact that even if two descriptors are unrelated, their matching score may contribute to the overall similarity. We formalise this problem and propose two related solutions, both aimed at equalising the individual contributions of the local descriptors in the final representation. These methods modify the aggregation stage by including a set of per-descriptor weights. They differ by the objective function that is optimised to compute those weights. The first is a "democratisation" strategy that aims at equalising the relative importance of each descriptor in the set comparison metric. The second one involves equalising the match of a single descriptor to the aggregated vector. These concurrent methods give a substantial performance boost over the state of the art in image search with short or mid-size vectors, as demonstrated by our experiments on standard public image retrieval benchmarks.

Albert Gordo, Jon Almazan, Naila Murray et al.

The goal of this work is to bring semantics into the tasks of text recognition and retrieval in natural images. Although text recognition and retrieval have received a lot of attention in recent years, previous works have focused on recognizing or retrieving exactly the same word used as a query, without taking the semantics into consideration. In this paper, we ask the following question: \emph{can we predict semantic concepts directly from a word image, without explicitly trying to transcribe the word image or its characters at any point?} For this goal we propose a convolutional neural network (CNN) with a weighted ranking loss objective that ensures that the concepts relevant to the query image are ranked ahead of those that are not relevant. This can also be interpreted as learning a Euclidean space where word images and concepts are jointly embedded. This model is learned in an end-to-end manner, from image pixels to semantic concepts, using a dataset of synthetically generated word images and concepts mined from a lexical database (WordNet). Our results show that, despite the complexity of the task, word images and concepts can indeed be associated with a high degree of accuracy

Luca Marchesotti, Naila Murray, Florent Perronnin

Aesthetic image analysis is the study and assessment of the aesthetic properties of images. Current computational approaches to aesthetic image analysis either provide accurate or interpretable results. To obtain both accuracy and interpretability by humans, we advocate the use of learned and nameable visual attributes as mid-level features. For this purpose, we propose to discover and learn the visual appearance of attributes automatically, using a recently introduced database, called AVA, which contains more than 250,000 images together with their aesthetic scores and textual comments given by photography enthusiasts. We provide a detailed analysis of these annotations as well as the context in which they were given. We then describe how these three key components of AVA - images, scores, and comments - can be effectively leveraged to learn visual attributes. Lastly, we show that these learned attributes can be successfully used in three applications: aesthetic quality prediction, image tagging and retrieval.

Naila Murray, Florent Perronnin

State-of-the-art patch-based image representations involve a pooling operation that aggregates statistics computed from local descriptors. Standard pooling operations include sum- and max-pooling. Sum-pooling lacks discriminability because the resulting representation is strongly influenced by frequent yet often uninformative descriptors, but only weakly influenced by rare yet potentially highly-informative ones. Max-pooling equalizes the influence of frequent and rare descriptors but is only applicable to representations that rely on count statistics, such as the bag-of-visual-words (BOV) and its soft- and sparse-coding extensions. We propose a novel pooling mechanism that achieves the same effect as max-pooling but is applicable beyond the BOV and especially to the state-of-the-art Fisher Vector -- hence the name Generalized Max Pooling (GMP). It involves equalizing the similarity between each patch and the pooled representation, which is shown to be equivalent to re-weighting the per-patch statistics. We show on five public image classification benchmarks that the proposed GMP can lead to significant performance gains with respect to heuristic alternatives.