Visual Representation Learning

Diffusion models have significantly reshaped the field of generative artificial intelligence and are now increasingly explored for their capacity in discriminative representation learning. Diffusion Transformer (DiT) has recently gained attention as a promising alternative to conventional U-Net-based diffusion models, demonstrating a promising avenue for downstream discriminative tasks via generative pre-training. However, its current training efficiency and representational capacity remain largely constrained due to the inadequate timestep searching and insufficient exploitation of DiT-specific feature representations. In light of this view, we introduce Automatically Selected Timestep (A-SelecT) that dynamically pinpoints DiT's most information-rich timestep from the selected transformer feature in a single run, eliminating the need for both computationally intensive exhaustive timestep searching and suboptimal discriminative feature selection. Extensive experiments on classification and segmentation benchmarks demonstrate that DiT, empowered by A-SelecT, surpasses all prior diffusion-based attempts efficiently and effectively.

Gait silhouettes, which can be encoded into binary gait codes, are widely adopted to representing motion patterns of pedestrian. Recent approaches commonly leverage visual backbones to encode gait silhouettes, achieving successful performance. However, they primarily focus on continuous visual features, overlooking the discrete nature of binary silhouettes that inherently share a discrete encoding space with natural language. Large Language Models (LLMs) have demonstrated exceptional capability in extracting discriminative features from discrete sequences and modeling long-range dependencies, highlighting their potential to capture temporal motion patterns by identifying subtle variations. Motivated by these observations, we explore bridging binary gait silhouettes and natural language within a binary encoding space. However, the encoding spaces of text tokens and binary gait silhouettes remain misaligned, primarily due to differences in token frequency and density. To address this issue, we propose the Contour-Velocity Tokenizer, which encodes binary gait silhouettes while reshaping their distribution to better align with the text token space. We then establish a dual-branch framework termed Silhouette Language Model, which enhances visual silhouettes by integrating discrete linguistic embeddings derived from LLMs. Implemented on mainstream gait backbones, SilLang consistently improves state-of-the-art methods across SUSTech1K, GREW, and Gait3D.

Tabular data are central to biomedical research, from liquid biopsy and bulk and single-cell transcriptomics to electronic health records and phenotypic profiling. Unlike images or sequences, however, tabular datasets lack intrinsic spatial organization: features are treated as unordered dimensions, and their relationships must be inferred implicitly by the model. This limits the ability of vision architectures to exploit local structure and higher-order feature interactions in non-spatial biomedical data. Here we introduce Dynamic Feature Mapping (Dynomap), an end-to-end deep learning framework that learns a task-optimized spatial topology of features directly from data. Dynomap jointly optimizes feature placement and prediction through a fully differentiable rendering mechanism, without relying on heuristics, predefined groupings, or external priors. By transforming high-dimensional tabular vectors into learned feature maps, Dynomap enables vision-based models to operate effectively on unordered biomedical inputs. Across multiple clinical and biological datasets, Dynomap consistently outperformed classical machine learning, modern deep tabular models, and existing vector-to-image approaches. In liquid biopsy data, Dynomap organized clinically relevant gene signatures into coherent spatial patterns and improved multiclass cancer subtype prediction accuracy by up to 18%. In a Parkinson disease voice dataset, it clustered disease-associated acoustic descriptors and improved accuracy by up to 8%. Similar gains and interpretable feature organization were observed in additional biomedical datasets. These results establish Dynomap as a general strategy for bridging tabular and vision-based deep learning and for uncovering structured, clinically relevant patterns in high-dimensional biomedical data.

Running AI models on smart edge devices can unlock versatile user experiences, but presents challenges due to limited compute and the need to handle multiple tasks simultaneously. This requires a vision encoder with small size but powerful and versatile representations. We present our method, Efficient Universal Perception Encoder (EUPE), which offers both inference efficiency and universally good representations for diverse downstream tasks. We achieve this by distilling from multiple domain-expert foundation vision encoders. Unlike previous agglomerative methods that directly scale down from multiple teachers to an efficient encoder, we demonstrate the importance of first scaling up to a large proxy teacher and then scaling down from this single teacher. Experiments show that EUPE achieves on-par or better performance than individual domain experts of the same size on diverse task domains and also outperforms previous agglomerative encoders. We release the full family of EUPE models and the code to foster future research.

Despite the remarkable success of large-scale pre-trained image representation models (i.e., vision encoders) across various vision tasks, they are predominantly trained on 2D image data and therefore often fail to capture 3D spatial relationships between objects and backgrounds in the real world, constraining their effectiveness in many downstream applications. To address this, we propose SpatialBoost, a scalable framework that enhances the spatial awareness of existing pre-trained vision encoders by injecting 3D spatial knowledge expressed in linguistic descriptions. The core idea involves converting dense 3D spatial information from 2D images into linguistic expressions, which is then used to inject such spatial knowledge into vision encoders through a Large Language Model (LLM). To this end, we adopt a multi-turn Chain-of-Thought (CoT) reasoning process that progressively incorporates dense spatial knowledge and builds hierarchical spatial understanding. To validate effectiveness, we adapt SpatialBoost to state-of-the-art vision encoders such as DINOv3, and evaluate its performance gains on a wide range of benchmarks requiring both 3D perception and general vision abilities. For instance, SpatialBoost improves DINOv3 performance from 55.9 to 59.7 mIoU on ADE20K, achieving state-of-the-art performance with 3.8% gain over the pre-trained DINOv3.

Generative models and vision encoders have largely advanced on separate tracks, optimized for different goals and grounded in different mathematical principles. Yet, they share a fundamental property: latent space Gaussianity. Generative models map Gaussian noise to images, while encoders map images to semantic embeddings whose coordinates empirically behave as Gaussian. We hypothesize that both are views of a shared latent source, the Universal Normal Embedding (UNE): an approximately Gaussian latent space from which encoder embeddings and DDIM-inverted noise arise as noisy linear projections. To test our hypothesis, we introduce NoiseZoo, a dataset of per-image latents comprising DDIM-inverted diffusion noise and matching encoder representations (CLIP, DINO). On CelebA, linear probes in both spaces yield strong, aligned attribute predictions, indicating that generative noise encodes meaningful semantics along linear directions. These directions further enable faithful, controllable edits (e.g., smile, gender, age) without architectural changes, where simple orthogonalization mitigates spurious entanglements. Taken together, our results provide empirical support for the UNE hypothesis and reveal a shared Gaussian-like latent geometry that concretely links encoding and generation. Code and data are available https://rbetser.github.io/UNE/

UAVs play an important role in applications such as autonomous exploration, disaster response, and infrastructure inspection. However, UAV VLN in complex 3D environments remains challenging. A key difficulty is the structural representation mismatch between 2D visual perception and the 3D trajectory decision space, which limits spatial reasoning. To this end, we propose SpatialFly, a geometry-guided spatial representation framework for UAV VLN. Operating on RGB observations without explicit 3D reconstruction, SpatialFly introduces a geometry-guided 2D adaptive representation mechanism. Specifically, the geometric prior injection module injects global structural cues into 2D semantic tokens to provide scene-level geometric guidance. The geometry-aware reparameterization module then uses geometry-conditioned cross-modal attention and gated residual fusion to adaptively reparameterize the visual tokens. Experimental results show that SpatialFly consistently outperforms state-of-the-art UAV VLN baselines across both seen and unseen environments, reducing NE by 4.03m and improving SR by 1.27% over the strongest baseline on the unseen Full split. Additional trajectory-level analysis shows that SpatialFly produces trajectories with better path alignment and smoother, more stable motion.

Modern vision models must capture image-level context without sacrificing local detail while remaining computationally affordable. We revisit this tradeoff and advance a simple principle: decouple the roles of global reasoning and local representation. To operationalize this principle, we introduce ConvNeur, a two-branch architecture in which a lightweight neural memory branch aggregates global context on a compact set of tokens, and a locality-preserving branch extracts fine structure. A learned gate lets global cues modulate local features without entangling their objectives. This separation yields subquadratic scaling with image size, retains inductive priors associated with local processing, and reduces overhead relative to fully global attention. On standard classification, detection, and segmentation benchmarks, ConvNeur matches or surpasses comparable alternatives at similar or lower compute and offers favorable accuracy versus latency trade-offs at similar budgets. These results support the view that efficiency follows global-local decoupling.

Tracking Any Point (TAP) has emerged as a fundamental tool for video understanding. Current approaches adapt Vision Foundation Models (VFMs) like DINOv2 via offline finetuning or test-time optimization. However, these VFMs rely on static image pre-training, which is inherently sub-optimal for capturing dense temporal correspondence in videos. To address this, we propose Mask-to-Point (M2P) learning, which leverages rich video object segmentation (VOS) mask annotations to improve VFMs for dense point tracking. Our M2P introduces three new mask-based constraints for weakly-supervised representation learning. First, we propose a local structure consistency loss, which leverages Procrustes analysis to model the cohesive motion of points lying within a local structure, achieving more reliable point-to-point matching learning. Second, we propose a mask label consistency (MLC) loss, which enforces that sampled foreground points strictly match foreground regions across frames. The proposed MLC loss can be regarded as a regularization, which stabilizes training and prevents convergence to trivial solutions. Finally, mask boundary constrain is applied to explicitly supervise boundary points. We show that our weaklysupervised M2P models significantly outperform baseline VFMs with efficient training by using only 3.6K VOS training videos. Notably, M2P achieves 12.8% and 14.6% performance gains over DINOv2-B/14 and DINOv3-B/16 on the TAP-Vid-DAVIS benchmark, respectively. Moreover, the proposed M2P models are used as pre-trained backbones for both test-time optimized and offline fine-tuned TAP tasks, demonstrating its potential to serve as general pre-trained models for point tracking. Code will be made publicly available upon acceptance.

Recent visual autonomous perception systems achieve remarkable performances with deep representation learning. However, they fail in scenarios with challenging illumination.While event cameras can mitigate this problem, there is a lack of a large-scale dataset to develop event-enhanced deep visual perception models in autonomous driving scenes. To address the gap, we present the eAP (event-enhanced Autonomous Perception) dataset, the largest dataset with event cameras for autonomous perception. We demonstrate how eAP can facilitate the study of different autonomous perception tasks, including 3D vehicle detection and object time-to-contact (TTC) estimation, through deep representation learning. Based on eAP, we demonstrate the ffrst successful use of events to improve a popular 3D vehicle detection network in challenging illumination scenarios. eAP also enables a devoted study of the representation learning problem of object TTC estimation. We show how a geometryaware representation learning framework leads to the best eventbased object TTC estimation network that operates at 200 FPS. The dataset, code, and pre-trained models will be made publicly available for future research.

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.

Training and evaluation in multi-channel imaging (MCI) remains challenging due to heterogeneous channel configurations arising from varying staining protocols, sensor types, and acquisition settings. This heterogeneity limits the applicability of fixed-channel encoders commonly used in general computer vision. Recent Multi-Channel Vision Transformers (MC-ViTs) address this by enabling flexible channel inputs, typically by jointly encoding patch tokens from all channels within a unified attention space. However, unrestricted token interactions across channels can lead to feature dilution, reducing the ability to preserve channel-specific semantics that are critical in MCI data. To address this, we propose Decoupled Vision Transformer (DC-ViT), which explicitly regulates information sharing using Decoupled Self-Attention (DSA), which decomposes token updates into two complementary pathways: spatial updates that model intra-channel structure, and channel-wise updates that adaptively integrate cross-channel information. This decoupling mitigates informational collapse while allowing selective inter-channel interaction. To further exploit these enhanced channel-specific representations, we introduce Decoupled Aggregation (DAG), which allows the model to learn task-specific channel importances. Extensive experiments across three MCI benchmarks demonstrate consistent improvements over existing MC-ViT approaches.

Visual navigation requires agents to reach goals in complex environments through perception and planning. World models address this task by simulating action-conditioned state transitions to predict future observations. Current navigation world models typically learn state evolution under actions within the compressed latent space of a Variational Autoencoder, where spatial compression often discards fine-grained structural information and hinders precise control. To better understand the propagation characteristics of different representations, we conduct a linear dynamics probe and observe that dense DINOv2 features exhibit stronger linear predictability for action-conditioned transitions. Motivated by this observation, we propose the Representation Autoencoder-based Navigation World Model (RAE-NWM), which models navigation dynamics in a dense visual representation space. We employ a Conditional Diffusion Transformer with Decoupled Diffusion Transformer head (CDiT-DH) to model continuous transitions, and introduce a separate time-driven gating module for dynamics conditioning to regulate action injection strength during generation. Extensive evaluations show that modeling sequential rollouts in this space improves structural stability and action accuracy, benefiting downstream planning and navigation.

Recent advances in multimodal large language models largely rely on CLIP-based visual encoders, which emphasize global semantic alignment but struggle with fine-grained visual understanding. In contrast, DINOv3 provides strong pixel-level perception yet lacks coarse-grained semantic abstraction, leading to limited multi-granularity reasoning. To address this gap, we propose Granulon, a novel DINOv3-based MLLM with adaptive granularity augmentation. Granulon introduces a text-conditioned granularity Controller that dynamically adjusts the visual abstraction level according to the semantic scope of the textual input, and an Adaptive Token Aggregation module that performs granularity-guided pooling and relation-aware clustering to produce compact, semantically rich visual tokens. This design enables unified "pixel-to-fine-to-coarse" reasoning within a single forward pass. Extensive and interpretable experiments demonstrate that Granulon improves accuracy by ~30% and reduces hallucination by ~20%, outperforming all visual encoders under identical settings.

Visual Document Retrieval (VDR) requires representations that capture both fine-grained visual details and global document structure to ensure retrieval efficacy while maintaining computational efficiency. Existing VDR models struggle to balance effectiveness and efficiency when processing high-resolution documents: they often either lose fine-grained information or generate an excessive number of visual tokens, resulting in significant indexing overhead and high retrieval latency. In this work, we rethink the visual encoding mechanism and propose a new X-VisEmb paradigm that progresses from multi-resolution sampling and encoding, through cross-granularity feature fusion, to adaptive representation distillation. A preliminary study validates its feasibility and effectiveness in capturing complementary visual cues at varying scales. Building on the insights, we develop MURE, a novel framework that employs VLMs as a hierarchical multi-resolution encoder, integrates resolution-level Matryoshka representation learning (RMRL) for effective feature fusion, and applies a semantic-aware hierarchical clustering mechanism for visual token compression. Experiments on two widely used VDR benchmarks show that our MURE framework consistently beats strong baselines. Furthermore, it significantly outperforms ColPali with only 50% of its visual token budget.

The limited understanding capacity of the visual encoder in Contrastive Language-Image Pre-training (CLIP) has become a key bottleneck for downstream performance. This capacity includes both Discriminative Ability (D-Ability), which reflects class separability, and Detail Perceptual Ability (P-Ability), which focuses on fine-grained visual cues. Recent solutions use diffusion models to enhance representations by conditioning image reconstruction on CLIP visual tokens. We argue that such paradigms may compromise D-Ability and therefore fail to effectively address CLIP's representation limitations. To address this, we integrate contrastive signals into diffusion-based reconstruction to pursue more comprehensive visual representations. We begin with a straightforward design that augments the diffusion process with contrastive learning on input images. However, empirical results show that the naive combination suffers from gradient conflict and yields suboptimal performance. To balance the optimization, we introduce the Diffusion Contrastive Reconstruction (DCR), which unifies the learning objective. The key idea is to inject contrastive signals derived from each reconstructed image, rather than from the original input, into the diffusion process. Our theoretical analysis shows that the DCR loss can jointly optimize D-Ability and P-Ability. Extensive experiments across various benchmarks and multi-modal large language models validate the effectiveness of our method. The code is available at https://github.com/boyuh/DCR.

Large vision-language models (LVLMs) demonstrate strong performance on diagram understanding benchmarks, yet they still struggle with understanding relationships between elements, particularly those represented by nodes and directed edges (e.g., arrows and lines). To investigate the underlying causes of this limitation, we probe the internal representation of LVLMs using a carefully constructed synthetic diagram dataset based on directed graphs. Our probing experiments reveal that edge information is not linearly separable in the vision encoder and becomes linearly encoded only in the text tokens in the language model. In contrast, node information and global structural features are already linearly encoded in individual hidden states of the vision encoder. These findings suggest that the stage at which linearly separable representations are formed varies depending on the type of visual information. In particular, the delayed emergence of edge representations may help explain why LVLMs struggle with relational understanding, such as interpreting edge directions, which require more abstract, compositionally integrated processes.

Unifying text-image contrastive learning and text-to-image (T2I) generation in a single end-to-end model is challenging because the two objectives demand opposing masking regimes: contrastive alignment needs near-complete visible tokens, while masked generative modeling needs heavy corruption. We introduce DREAM, a unified framework that resolves this conflict through Masking Warmup, a schedule that shifts the center of the masking distribution over training, so low and high masking ratios coexist at every step. This co-exposure lets a single jointly-trained encoder serve both objectives. The resulting stable optimization unlocks Semantically Aligned Decoding at inference: the text encoder, trained against visual embeddings at all masking ratios, can score partially generated images and select the best trajectory with as little as 12.5% of the image decoded, improving both FID and throughput. DREAM outperforms its single-objective baselines, CLIP and FLUID: on ImageNet linear-probing (+1.1%), 5-shot transfer (+4.1%), ADE20K segmentation (+1.9%), and NYU depth estimation (+6.25%) over CLIP, and on CC12M FID (+6.2%) over FLUID while maintaining CLIP Score. Together, these gains show that text-image contrastive and generative objectives, when properly unified, are synergistic rather than competing.

World models compress rich sensory streams into compact latent codes that anticipate future observations. We let separate agents acquire such models from distinct viewpoints of the same environment without any parameter sharing or coordination. After training, their internal representations exhibit a striking emergent property: the two latent spaces are related by an approximate linear isometry, enabling transparent translation between them. This geometric consensus survives large viewpoint shifts and scant overlap in raw pixels. Leveraging the learned alignment, a classifier trained on one agent can be ported to the other with no additional gradient steps, while distillation-like migration accelerates later learning and markedly reduces total compute. The findings reveal that predictive learning objectives impose strong regularities on representation geometry, suggesting a lightweight path to interoperability among decentralized vision systems. The code is available at https://anonymous.4open.science/r/Social-JEPA-5C57.

Learning efficient and expressive visual representation has long been the pursuit of computer vision research. While Vision Transformers (ViTs) gradually replace traditional Convolutional Neural Networks (CNNs) as more scalable vision learners, their applications are plagued by the quadratic complexity of the self-attention mechanism. To address the challenge, we introduce a new linear-time sequence modeling method Test-Time Training (TTT) into vision and propose Vision-TTT, which treats visual sequences as datasets and compresses the visual token sequences in a novel self-supervised learning manner. By incorporating the dual-dataset strategy and Conv2d-based dataset preprocessing, Vision-TTT effectively extends vanilla TTT to model 2D visual correlations with global receptive fields. Extensive experiments show that Vittt-T/S/B achieve $77.7\%,81.8\%,82.7\%$ Top-1 accuracy on ImageNet classification and also greatly outperform their counterparts on downstream tasks. At $1280\times1280$ resolution, Vittt-T reduces FLOPs by $79.4\%$ and runs $4.72\times$ faster with $88.9\%$ less memory than DeiT-T. These results demonstrate the expressiveness and efficiency of Vision-TTT as a strong candidate for the next-generation generic visual backbone.

Compositional generalization, the ability to recognize familiar parts in novel contexts, is a defining property of intelligent systems. Although modern models are trained on massive datasets, they still cover only a tiny fraction of the combinatorial space of possible inputs, raising the question of what structure representations must have to support generalization to unseen combinations. We formalize three desiderata for compositional generalization under standard training (divisibility, transferability, stability) and show they impose necessary geometric constraints: representations must decompose linearly into per-concept components, and these components must be orthogonal across concepts. This provides theoretical grounding for the Linear Representation Hypothesis: the linear structure widely observed in neural representations is a necessary consequence of compositional generalization. We further derive dimension bounds linking the number of composable concepts to the embedding geometry. Empirically, we evaluate these predictions across modern vision models (CLIP, SigLIP, DINO) and find that representations exhibit partial linear factorization with low-rank, near-orthogonal per-concept factors, and that the degree of this structure correlates with compositional generalization on unseen combinations. As models continue to scale, these conditions predict the representational geometry they may converge to. Code is available at https://github.com/oshapio/necessary-compositionality.

Vision disorders significantly impact millions of lives, altering how visual information is processed and perceived. In this work, a computational framework was developed using the BrokenEyes system to simulate five common eye disorders: Age-related macular degeneration, cataract, glaucoma, refractive errors, and diabetic retinopathy and analyze their effects on neural-like feature representations in deep learning models. Leveraging a combination of human and non-human datasets, models trained under normal and disorder-specific conditions revealed critical disruptions in feature maps, particularly for cataract and glaucoma, which align with known neural processing challenges in these conditions. Evaluation metrics such as activation energy and cosine similarity quantified the severity of these distortions, providing insights into the interplay between degraded visual inputs and learned representations.

Vision-language foundation models (VLFMs) promise zero-shot and retrieval understanding for Earth observation. While operational satellite systems often lack full multi-spectral coverage, making RGB-only inference highly desirable for scalable deployment, the adoption of VLFMs for satellite imagery remains hindered by two factors: (1) multi-spectral inputs are informative but difficult to exploit consistently due to band redundancy and misalignment; and (2) CLIP-style text encoders limit semantic expressiveness and weaken fine-grained alignment. We present SATtxt, a spectrum-aware VLFM that operates with RGB inputs only at inference while retaining spectral cues learned during training. Our framework comprises two stages. First, Spectral Representation Distillation transfers spectral priors from a frozen multi-spectral teacher to an RGB student via a lightweight projector. Second, Spectrally Grounded Alignment with Instruction-Augmented LLMs bridges the distilled visual space and an expressive LLM embedding space. Across EuroSAT, BigEarthNet, and ForestNet, SATtxt improves zero-shot classification on average by 4.2%, retrieval by 5.9%, and linear probing by 2.7% over baselines, showing an efficient path toward spectrum-aware vision-language learning for Earth observation. Project page: https://ikhado.github.io/sattxt/

Vision Transformers (ViTs), when pre-trained on large-scale data, provide general-purpose representations for diverse downstream tasks. However, artifacts in ViTs are widely observed across different supervision paradigms and downstream tasks. Through systematic analysis of artifacts in ViTs, we find that their fundamental mechanisms have yet to be sufficiently elucidated. In this paper, through systematic analysis, we conclude that these artifacts originate from a lazy aggregation behavior: ViT uses semantically irrelevant background patches as shortcuts to represent global semantics, driven by global attention and Coarse-grained semantic supervision. Our solution selectively integrates patch features into the CLS token, reducing the influence of background-dominated shortcuts and consistently improving performance across 12 benchmarks under label-, text-, and self-supervision. We hope this work offers a new perspective on ViT behavior.

What does it mean for a visual system to truly understand affordance? We argue that this understanding hinges on two complementary capacities: geometric perception, which identifies the structural parts of objects that enable interaction, and interaction perception, which models how an agent's actions engage with those parts. To test this hypothesis, we conduct a systematic probing of Visual Foundation Models (VFMs). We find that models like DINO inherently encode part-level geometric structures, while generative models like Flux contain rich, verb-conditioned spatial attention maps that serve as implicit interaction priors. Crucially, we demonstrate that these two dimensions are not merely correlated but are composable elements of affordance. By simply fusing DINO's geometric prototypes with Flux's interaction maps in a training-free and zero-shot manner, we achieve affordance estimation competitive with weakly-supervised methods. This final fusion experiment confirms that geometric and interaction perception are the fundamental building blocks of affordance understanding in VFMs, providing a mechanistic account of how perception grounds action.

Latent action representations learned from unlabeled videos have recently emerged as a promising paradigm for pretraining vision-language-action (VLA) models without explicit robot action supervision. However, latent actions derived solely from RGB observations primarily encode appearance-driven dynamics and lack explicit 3D geometric structure, which is essential for precise and contact-rich manipulation. To address this limitation, we introduce UniLACT, a transformer-based VLA model that incorporates geometric structure through depth-aware latent pretraining, enabling downstream policies to inherit stronger spatial priors. To facilitate this process, we propose UniLARN, a unified latent action learning framework based on inverse and forward dynamics objectives that learns a shared embedding space for RGB and depth while explicitly modeling their cross-modal interactions. This formulation produces modality-specific and unified latent action representations that serve as pseudo-labels for the depth-aware pretraining of UniLACT. Extensive experiments in both simulation and real-world settings demonstrate the effectiveness of depth-aware unified latent action representations. UniLACT consistently outperforms RGB-based latent action baselines under in-domain and out-of-domain pretraining regimes, as well as on both seen and unseen manipulation tasks.The project page is at https://manishgovind.github.io/unilact-vla/

Foundation models are transforming Earth Observation (EO), yet the diversity of EO sensors and modalities makes a single universal model unrealistic. Multiple specialized EO foundation models (EOFMs) will likely coexist, making efficient knowledge transfer across modalities essential. Most existing EO pretraining relies on masked image modeling, which emphasizes local reconstruction but provides limited control over global semantic structure. To address this, we propose a dual-teacher contrastive distillation framework for multispectral imagery that aligns the student's pretraining objective with the contrastive self-distillation paradigm of modern optical vision foundation models (VFMs). Our approach combines a multispectral teacher with an optical VFM teacher, enabling coherent cross-modal representation learning. Experiments across diverse optical and multispectral benchmarks show that our model adapts to multispectral data without compromising performance on optical-only inputs, achieving state-of-the-art results in both settings, with an average improvement of 3.64 percentage points in semantic segmentation, 1.2 in change detection, and 1.31 in classification tasks. This demonstrates that contrastive distillation provides a principled and efficient approach to scalable representation learning across heterogeneous EO data sources. Project page: \textcolor{magenta}{https://wolfilip.github.io/DEO/}.

World models learned from high-dimensional visual observations allow agents to make decisions and plan directly in latent space, avoiding pixel-level reconstruction. However, recent latent predictive architectures (JEPAs), including the DINO world model (DINO-WM), display a degradation in test time robustness due to their sensitivity to ``slow features". These include visual variations such as background changes and distractors that are irrelevant to the task being solved. We address this limitation by augmenting the predictive objective with a bisimulation encoder that enforces control-relevant state equivalence, mapping states with similar transition dynamics to nearby latent states while limiting contributions from slow features. We evaluate our model on a navigation task (PointMaze) and on a manipulation task (PushT) under different test-time background changes and visual distractors. Across all benchmarks, our model consistently improves robustness to slow features while operating in a reduced latent space, up to $10\times$ smaller than that of DINO-WM. Moreover, our model is agnostic to the choice of pre-trained visual encoder and maintains robustness when paired with DINOv2, SimDINOv2, and iBOT features.

Modern vision models increasingly rely on rich semantic representations that extend beyond class labels to include descriptive concepts, attributes, and contextual cues. However, semantic concepts are not uniformly represented across classes: a concept may be frequent globally yet remain underrepresented within a class, resulting in low class-concept coverage. We formalize this class-concept phenomenon as Semantic Coverage Imbalance (SCI) and study its relationship with learned semantic representations. We find that lower class-concept coverage is consistently associated with weaker semantic representations. To address this, we introduce Coverage-aware Semantic Representation Learning, which shares more semantic structure for low coverage or uncertain relations while preserving more pair-specific structure for high coverage relations. We instantiate this principle in SemCovNet with Coverage-Calibrated Semantic Sharing (CCSS), a coverage- and uncertainty-aware partial-pooling mechanism. Using a range of vision datasets spanning facial attributes and fine-grained recognition, as well as real-world medical datasets, we show that SCI is widespread, degrades representation quality, and can be mitigated through coverage-aware semantic representation learning.

Compositional generalization, the ability to reason about novel combinations of familiar concepts, is fundamental to human cognition and a critical challenge for machine learning. Object-centric (OC) representations, which encode a scene as a set of objects, are often argued to support such generalization, but systematic evidence in visually rich settings is limited. We introduce a Visual Question Answering benchmark across three controlled visual worlds (CLEVRTex, Super-CLEVR, and MOVi-C) to measure how well vision encoders, with and without object-centric biases, generalize to unseen combinations of object properties. To ensure a fair and comprehensive comparison, we carefully account for training data diversity, sample size, representation size, downstream model capacity, and compute. We use DINOv2 and SigLIP2, two widely used vision encoders, as the foundation models and their OC counterparts. Our key findings reveal that (1) OC approaches are superior in harder compositional generalization settings; (2) original dense representations surpass OC only on easier settings and typically require substantially more downstream compute; and (3) OC models are more sample efficient, achieving stronger generalization with fewer images, whereas dense encoders catch up or surpass them only with sufficient data and diversity. Overall, object-centric representations offer stronger compositional generalization when any one of dataset size, training data diversity, or downstream compute is constrained.