Abstract:Building general-purpose models that can effectively perceive the world through multimodal signals has been a long-standing goal. Current approaches involve integrating separately pre-trained components, such as connecting vision encoders to LLMs and continuing multimodal training. While such approaches exhibit remarkable sample efficiency, it remains an open question whether such late-fusion architectures are inherently superior. In this work, we revisit the architectural design of native multimodal models (NMMs)--those trained from the ground up on all modalities--and conduct an extensive scaling laws study, spanning 457 trained models with different architectures and training mixtures. Our investigation reveals no inherent advantage to late-fusion architectures over early-fusion ones, which do not rely on image encoders. On the contrary, early-fusion exhibits stronger performance at lower parameter counts, is more efficient to train, and is easier to deploy. Motivated by the strong performance of the early-fusion architectures, we show that incorporating Mixture of Experts (MoEs) allows for models that learn modality-specific weights, significantly enhancing performance.
Abstract:We explore the potential of large-scale generative video models for autonomous driving, introducing an open-source auto-regressive video model (VaViM) and its companion video-action model (VaVAM) to investigate how video pre-training transfers to real-world driving. VaViM is a simple auto-regressive video model that predicts frames using spatio-temporal token sequences. We show that it captures the semantics and dynamics of driving scenes. VaVAM, the video-action model, leverages the learned representations of VaViM to generate driving trajectories through imitation learning. Together, the models form a complete perception-to-action pipeline. We evaluate our models in open- and closed-loop driving scenarios, revealing that video-based pre-training holds promise for autonomous driving. Key insights include the semantic richness of the learned representations, the benefits of scaling for video synthesis, and the complex relationship between model size, data, and safety metrics in closed-loop evaluations. We release code and model weights at https://github.com/valeoai/VideoActionModel
Abstract:Understanding the 3D geometry and semantics of driving scenes is critical for developing of safe autonomous vehicles. While 3D occupancy models are typically trained using voxel-based supervision with standard losses (e.g., cross-entropy, Lovasz, dice), these approaches treat voxel predictions independently, neglecting their spatial relationships. In this paper, we propose GaussRender, a plug-and-play 3D-to-2D reprojection loss that enhances voxel-based supervision. Our method projects 3D voxel representations into arbitrary 2D perspectives and leverages Gaussian splatting as an efficient, differentiable rendering proxy of voxels, introducing spatial dependencies across projected elements. This approach improves semantic and geometric consistency, handles occlusions more efficiently, and requires no architectural modifications. Extensive experiments on multiple benchmarks (SurroundOcc-nuScenes, Occ3D-nuScenes, SSCBench-KITTI360) demonstrate consistent performance gains across various 3D occupancy models (TPVFormer, SurroundOcc, Symphonies), highlighting the robustness and versatility of our framework. The code is available at https://github.com/valeoai/GaussRender.
Abstract:Existing vehicle trajectory prediction models struggle with generalizability, prediction uncertainties, and handling complex interactions. It is often due to limitations like complex architectures customized for a specific dataset and inefficient multimodal handling. We propose Perceiver with Register queries (PerReg+), a novel trajectory prediction framework that introduces: (1) Dual-Level Representation Learning via Self-Distillation (SD) and Masked Reconstruction (MR), capturing global context and fine-grained details. Additionally, our approach of reconstructing segmentlevel trajectories and lane segments from masked inputs with query drop, enables effective use of contextual information and improves generalization; (2) Enhanced Multimodality using register-based queries and pretraining, eliminating the need for clustering and suppression; and (3) Adaptive Prompt Tuning during fine-tuning, freezing the main architecture and optimizing a small number of prompts for efficient adaptation. PerReg+ sets a new state-of-the-art performance on nuScenes [1], Argoverse 2 [2], and Waymo Open Motion Dataset (WOMD) [3]. Remarkable, our pretrained model reduces the error by 6.8% on smaller datasets, and multi-dataset training enhances generalization. In cross-domain tests, PerReg+ reduces B-FDE by 11.8% compared to its non-pretrained variant.
Abstract:Multimodal LLMs have reached remarkable levels of proficiency in understanding multimodal inputs, driving extensive research to develop increasingly powerful models. However, much less attention has been paid to understanding and explaining the underlying mechanisms of these models. Most existing explainability research examines these models only in their final states, overlooking the dynamic representational shifts that occur during training. In this work, we systematically analyze the evolution of hidden state representations to reveal how fine-tuning alters the internal structure of a model to specialize in new multimodal tasks. Using a concept-based approach, we map hidden states to interpretable visual and textual concepts, enabling us to trace changes in encoded concepts across modalities as training progresses. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by shifting those in the original model. Finally, we explore the practical impact of our findings on model steering, showing that we can adjust multimodal LLMs behaviors without any training, such as modifying answer types, captions style, or biasing the model toward specific responses. Our work sheds light on how multimodal representations evolve through fine-tuning and offers a new perspective for interpreting model adaptation in multimodal tasks. The code for this project is publicly available at https://github.com/mshukor/xl-vlms.
Abstract:Motion forecasting (MF) for autonomous driving aims at anticipating trajectories of surrounding agents in complex urban scenarios. In this work, we investigate a mixed strategy in MF training that first pre-train motion forecasters on pseudo-labeled data, then fine-tune them on annotated data. To obtain pseudo-labeled trajectories, we propose a simple pipeline that leverages off-the-shelf single-frame 3D object detectors and non-learning trackers. The whole pre-training strategy including pseudo-labeling is coined as PPT. Our extensive experiments demonstrate that: (1) combining PPT with supervised fine-tuning on annotated data achieves superior performance on diverse testbeds, especially under annotation-efficient regimes, (2) scaling up to multiple datasets improves the previous state-of-the-art and (3) PPT helps enhance cross-dataset generalization. Our findings showcase PPT as a promising pre-training solution for robust motion forecasting in diverse autonomous driving contexts.
Abstract:Understanding deep models is crucial for deploying them in safety-critical applications. We introduce GIFT, a framework for deriving post-hoc, global, interpretable, and faithful textual explanations for vision classifiers. GIFT starts from local faithful visual counterfactual explanations and employs (vision) language models to translate those into global textual explanations. Crucially, GIFT provides a verification stage measuring the causal effect of the proposed explanations on the classifier decision. Through experiments across diverse datasets, including CLEVR, CelebA, and BDD, we demonstrate that GIFT effectively reveals meaningful insights, uncovering tasks, concepts, and biases used by deep vision classifiers. Our code, data, and models are released at https://github.com/valeoai/GIFT.
Abstract:We consider the problem of text-to-video generation tasks with precise control for various applications such as camera movement control and video-to-video editing. Most methods tacking this problem rely on providing user-defined controls, such as binary masks or camera movement embeddings. In our approach we propose OnlyFlow, an approach leveraging the optical flow firstly extracted from an input video to condition the motion of generated videos. Using a text prompt and an input video, OnlyFlow allows the user to generate videos that respect the motion of the input video as well as the text prompt. This is implemented through an optical flow estimation model applied on the input video, which is then fed to a trainable optical flow encoder. The output feature maps are then injected into the text-to-video backbone model. We perform quantitative, qualitative and user preference studies to show that OnlyFlow positively compares to state-of-the-art methods on a wide range of tasks, even though OnlyFlow was not specifically trained for such tasks. OnlyFlow thus constitutes a versatile, lightweight yet efficient method for controlling motion in text-to-video generation. Models and code will be made available on GitHub and HuggingFace.
Abstract:Large Language Models (LLMs) have demonstrated remarkable success in both textual and multimodal domains. However, this success often comes with substantial computational costs, particularly when handling lengthy sequences of multimodal inputs. This has sparked many efforts focusing on enhancing efficiency during training and inference. In this study, we investigate the computation redundancy in Multimodal Large Language Models (MLLMs) during inference. We propose different methods to skip computations, such as skipping entire blocks, FFN or self-attention (SA) layers. Additionally, we explore parallelizing certain layers, such as FFN and SA layers. Our findings validate that (1) significant amount of computations can be avoided at inference time, especially for tasks such as Visual Question Answering (VQA). (2) Skipping computations during training can recover 97% of the original performance, even when skipping half of the blocks or removing 70% of the weights. Alternatively, (3) properly training with smaller LLMs can yield comparable performance to LLMs 2 or 3 times larger. To conclude, we extend our investigation to recent MLLMs, such as LLaVA-1.5, showing similar observations. Our work show that there is redundant computations inside MLLMs and thus the potential for significantly improving inference costs without sacrificing performance. The code is available here: https://github.com/mshukor/ima-lmms.
Abstract:Vision Language Models (VLMs) have shown impressive performances on numerous tasks but their zero-shot capabilities can be limited compared to dedicated or fine-tuned models. Yet, fine-tuning VLMs comes with limitations as it requires `white-box' access to the model's architecture and weights as well as expertise to design the fine-tuning objectives and optimize the hyper-parameters, which are specific to each VLM and downstream task. In this work, we propose LLM-wrapper, a novel approach to adapt VLMs in a `black-box' manner by leveraging large language models (LLMs) so as to reason on their outputs. We demonstrate the effectiveness of LLM-wrapper on Referring Expression Comprehension (REC), a challenging open-vocabulary task that requires spatial and semantic reasoning. Our approach significantly boosts the performance of off-the-shelf models, resulting in competitive results when compared with classic fine-tuning.