Latent Space

Text-to-motion generation must produce motions that are semantically correct, temporally coherent, and physically plausible. A natural approach is to first project motion data into a structured semantic space and then train a generative model within that space. Such a paradigm has been highly successful in image generation through Representation Autoencoders (RAEs), where a frozen self-supervised encoder provides semantic features for diffusion or flow models to learn from. However, direct transfer of such a paradigm to motion space using Motion-JEPA as the frozen encoder fails dramatically. We diagnose this failure geometrically and identify two motion-specific bottlenecks: (1) the JEPA feature space is spectrally ill-conditioned, making the Gaussian-to-data transport unstable; and (2) even with a well-conditioned spectrum, flow residuals tend to align with decoder-sensitive directions, where small latent errors are amplified into large motion artifacts after decoding. Based on these insights, we propose MoRAE. MoRAE addresses the two bottlenecks separately. A compact bottleneck distills the structured JEPA representation while removing weak and redundant directions, bringing the latent spectrum into a transport-stable regime. Motion-coupled training then aligns the retained latent geometry with the decoder, making characteristic flow errors less costly after decoding. With this flow-friendly latent, a standard non-autoregressive Flow-Matching DiT achieves state-of-the-art performance.

Differentially private (DP) training of text-conditioned generative models suffers a utility cliff at strong privacy. We revisit this problem through the geometry of rectified flows: along the straight interpolation between noise and data, the Bayes-optimal velocity is governed to leading order at the noise end by a few class-conditional moments, and increasingly sample-specific structure matters toward the data end. StraightDP exploits this heterogeneity end to end. A small budget share releases whitened class-conditional moments once, to be distilled into the weights or injected at sampling time. The rest is spent by pre-declared DP-SGD toward the data end, beyond the moments' reach. At $\varepsilon=1$ on MNIST, the released moments alone already attain $0.76$ downstream accuracy with prototype-like samples and an FID of $237$, and uniform DP-SGD attains $0.21$. The pipeline built on the release reaches $0.81$ accuracy at FID $56$ in a public latent space. Constraining per-token stream norms of the multimodal backbone leaves the pretraining loss unchanged yet improves downstream accuracy in the extreme-noise pixel-space regime, and its accuracy effect becomes monotonically more favorable as privacy strengthens. The released moments also port to frozen SD3-medium, where sampling-time injection beats DP-LoRA training at a fraction of the budget.

Fine-grained segmentation of auricular structures in CT is challenging because the ear occupies a small image region, cartilage boundaries are highly irregular, and interfaces between cartilage and surrounding soft tissues are often ambiguous. Clinical annotations may also include both composite structures containing cartilage and adjacent skin and their corresponding cartilage-only regions, producing nested and overlapping labels. We propose a world-model-based segmentation framework that enables iterative anatomical reasoning beyond conventional feed-forward prediction. Built on an encoder-decoder architecture, the framework introduces a deterministic recurrent state-space model into the intermediate latent space. Multi-scale encoder features and partially decoded representations are fused to form a structural observation that initializes the latent dynamics. During inference, the model performs a three-step latent rollout without ground-truth guidance. Hierarchical anatomical actions update the recurrent state and progressively refine the latent representation. The resulting latent trajectory is projected back into the decoder and combined with high-resolution features to produce the final segmentation. To learn reliable latent transitions, we introduce a balanced hierarchical action objective that addresses foreground sparsity, missing anatomical groups, and imbalance between add and remove operations. Extensive experiments show that the proposed framework consistently improves segmentation accuracy and reduces HD95 by more than 43% for small, irregular, and overlapping auricular structures in CT. These results demonstrate the effectiveness of latent world-model reasoning for challenging medical image segmentation.

Pathology foundation models (FMs) are models trained on vast amounts of typically unlabeled data and have been shown to yield regularized latent spaces that can be used effectively in downstream classification tasks. This is also true for the classification of mitotic figures vs. other cells. However, it is so far unclear if the latent space of current FMs provides features that are discriminant and spatially suitably resolved to also serve as a backbone for dense object detection paradigms. In this work, we investigate this question for common current pathology FMs (UNI, UNI2-h, Virchow, Virchow2, H-optimus-0, H-optimus-1) and compare their performance against a fully end-to-end trained baseline based on a ResNet50 architecture. We combine FM backbones with representatives of single stage, dual stage and self-attention-based detectors (RetinaNet, Faster R-CNN, Deformable DETR respectively) on the multi-domain MIDOG++ dataset, and on the TUPAC16 dataset as an out-of-domain case. We show that the H-optimus-0 and Virchow models yielded competitive performance, indicating that the latent spaces of current FMs, all trained on image-level self-supervision, are suitable for direct mitotic figure detection and may be slightly more robust on our out-of-domain test case. All code is made available publicly at https://github.com/DeepMicroscopy/FM4MFdet.

Grounding human motion in language, and language in motion, is a central step toward physical AI systems that can understand, generate, and communicate human behavior. Unified motion--language systems first coupled the two directions through a shared discrete motion codebook, but quantization limits generation quality. The strongest generators buy quality back at growing cost: stacked residual codebooks enlarge the representation; masked decoding stages, long autoregressive rollouts, and denoising chains of tens to hundreds of steps stretch inference; even the continuous-latent designs among them reach their latent only through an iterative diffusion head; and none of this decoding machinery serves understanding. We therefore propose MUGEN, a unified motion--language framework that pays neither cost: no codebook, one draw. A single adaptive-length autoencoder compresses any-length motion into a few continuous latent slots, the system's only motion representation: the language model generates them for text-to-motion and reads them back for motion understanding. Depth-routed hidden states let each slot read from the transformer depth it needs, and a calibrated head predicts a joint distribution over the full latent set, so a single draw carries the text-conditional, cross-slot variation a description permits. At a decoding cost of K language-model steps, one draw, and one decoder pass, MUGEN leads language-model baselines on FID on HumanML3D while raising retrieval precision above the real-motion reference under the standard evaluator, achieves the best CIDEr and BLEU@4 scores, and surpasses the discrete-token state of the art on every retrieval and alignment metric on SnapMoGen.

Despite the impressive visuomotor capabilities enabled by Vision-Language-Action (VLA) models, their performance often degrades on challenging and out-of-domain tasks. Recent test-time steering and scaling methods improve performance without extensive data collection and retraining, but action samples often remain concentrated around similar behaviors and therefore inherit correlated failure modes. Moreover, existing methods apply the same intervention strategy at every timestep, regardless of whether the base policy is already likely to succeed. To address these limitations, we introduce $RL^2$, an adaptive inference-time steering framework that leverages Reinforcement Learning on VLA Latents. First, we train a lightweight offline RL policy conditioned on expressive latents extracted from the VLA action expert and compose its flow velocity with that of the frozen VLA during inference. This compositional steering strategy combines the behavioral priors of large-scale imitation learning with the action diversity induced by offline RL beyond dominant demonstration modes. We further discover that inference-time steering follows fundamentally different scaling laws under success and failure states, revealing that action diversity is most beneficial when the base VLA is likely to fail, but can unnecessarily perturb already-accurate actions when success is likely. Building on this insight, $RL^2$ activates compositional steering only when failure is predicted. Across the SIMPLER and PolaRiS benchmarks, $RL^2$ improves success rates by up to +17.3% in out-of-domain settings, while ablations and scaling studies demonstrate the importance of latent representations and RL training. Finally, real-world experiments demonstrate that these gains transfer beyond simulation, establishing $RL^2$ as a practical and modular steering framework for VLA deployment.

Different research lines use the term world model in different ways, yet they share a common aim: to capture how the world evolves under action in a form that supports perception, simulation, and planning. Two prominent realizations are neural predictors that learn dynamics in continuous vector spaces, and hand-built physics engines that expose explicit state and physical laws. Neural predictors scale from data but leave the form of the dynamics implicit; physics engines are inspectable and editable but difficult to construct at scale. We introduce VisualPatchWorld (VPW), which represents world dynamics as code. VPW first selects a qualitative dynamical form with short active probes, then fits that form's free parameters from recorded state-action traces by minimizing multi-step prediction error. The resulting programs can be rolled forward like a simulator, inspected in source form, and used inside model-predictive control; image-derived scene graphs can supply the live state at replan time. Across comparisons with prior code-based world models, VPW attains 69.0% mean planning success and exceeds the strongest code baseline by 23.5 points. The largest gains arise when choosing the correct qualitative dynamics is essential. Under the same planner, the induced models approach ground-truth engine success on navigation and grasp-rich control; a residual gap remains for contact-rich pushing, and checking a shortlist of promising plans in the engine closes most of that gap. These results establish a practical route toward automatically constructed code world models that are useful for planning. Code is available at https://github.com/HKBU-KnowComp/VisualPatchWorld/.

High-fidelity 3D generative modeling increasingly relies on the latent diffusion paradigm, where the reconstruction quality of the underlying 3D VAE becomes a primary bottleneck. Existing approaches largely follow two paradigms: sparse voxel-based representations achieve strong reconstruction quality but incur significant memory and computational overhead, while set-based representations are compact and continuous yet typically lag in fidelity due to latent sparsity and excessive global smoothness. We propose MSVS-VAE, a hierarchical set-based VAE that closes this fidelity gap without sacrificing compactness. Our key idea is to progressively densify anchored VecSet latents via hierarchical point-shuffle upsampling, increasing spatial capacity for fine-grained geometry modeling. To efficiently decode from the densified hierarchy, we replace global cross-attention with AVS-Conv, a geometry-aware local aggregation operator operating within local neighborhoods rather than the exhaustive latent set. We further introduce multi-scale query decoding to fuse coarse-to-fine latent features, where coarse scales provide stable global context, and fine scales refine localized geometry, reducing artifacts from overly local receptive fields. Extensive experiments on Objaverse, ABO, and in-the-wild benchmarks demonstrate that MSVS-VAE consistently outperforms prior set-based and voxel-based VAEs, delivering approximately 10x faster decoding than prior set-based methods and approximately 10x higher compactness than voxel-based baselines.

Video temporal grounding (VTG) aims to localize the continuous video interval described by a natural-language query. However, current VLM-based methods typically produce this interval indirectly through two endpoint outputs, represented either as discrete timestamp tokens or continuous boundary coordinates. These formulations differ in how endpoints are encoded, but not in what is predicted: the event interval remains a derived object, while interval validity, duration, and interval-level similarity are handled only implicitly. We propose TimePLE, which reformulates VTG from endpoint prediction to interval-native grounding by predicting a single joint distribution over valid temporal intervals. TimePLE maps each interval to a point in a canonical position-duration square, where every support point corresponds to a valid span and neighboring points represent geometrically similar intervals. Given a video and query, the VLM generates a single latent <|TIMESPAN|> token whose hidden state is decoded into a joint interval distribution, refined through duration-aware coordinate correction, and converted into continuous boundaries. The same interval representation is used to encode input temporal anchors, aligning video-side temporal evidence with output-side span prediction. To reliably align the latent span representation with complete event intervals, we curate 90K-scale grounded samples and human-verify 3K-scale benchmark annotations. Experiments across four VTG benchmarks show that TimePLE consistently outperforms endpoint prediction baselines, achieving an average mIoU of 58.9, with clear gains on short-duration and medium-duration events.

The usefulness of a variational autoencoder (VAE) depends on two properties of its latent space that are hard to obtain together: high encoding capacity in the individual latent variables, and a low-dimensional, disentangled organization of those variables. Weakening the Kullback-Leibler regularization raises capacity but degrades disentanglement, while strengthening it prunes latent variables away entirely. We formulate VAE training as a soft-constrained optimization problem that addresses both. First, we impose an entropy-based constraint (EC) on individual latent variables, showing that the entropy of a latent code upper-bounds the mutual information it carries about the generative factors of the data. Second, we propose a weight-filter method that exploits the slack of the soft constraint to prune low-entropy dimensions during downstream training. On dSprites, the EC raises the aggregate latent-variable activation score by 43-62% over a vanilla VAE, attains the highest FactorVAE score among the \b{eta} \b{eta}-VAE variants (0.891 vs 0.847), and lowers reconstruction error by up to 38%. On MNIST, the weight filter reduces the latent dimensionality supplied to a downstream classifier from ten to two while holding accuracy above 90%, converging in 37% fewer epochs than the same procedure without the EC. We also find that low-entropy discrete factors tend to merge into a single latent variable, whereas high-entropy continuous factors are distributed across several.

In multiple-input multiple-output (MIMO) semantic communication, imperfect channel state information (CSI) and equalization mismatch can seriously degrade semantic reconstruction quality. To address this issue, we propose a unified restoration flow matching (RFM)-based framework for channel refinement and equalization correction. Specifically, the channel RFM (CRFM) module is developed to refine the coarse channel, thereby improving channel estimation accuracy. Based on the refined channel, the developed semantic RFM (SRFM) module is employed to correct the residual distortions in the post-equalization latent space. The key idea is to formulate the two cascaded inverse problems of channel estimation and equalization as the unified conditional restoration task, in which the learned conditional velocity field guides the perturbed distribution towards the target distribution. To enhance the robustness of these two modules under various distortion conditions, we develop a dual-anchor perturbation training strategy that jointly learns near-manifold refinement and large-error correction, and implement inference through a few-step deterministic ordinary differential equation (ODE) solver. Extensive experiments on MIMO channels and visual semantic transmission tasks demonstrate that the proposed scheme improves key metrics for channel estimation and semantic reconstruction quality. Moreover, compared with representative diffusion-based generative baselines, the proposed method requires fewer sampling steps.

In real-world machine learning applications, incomplete observations create a fundamental challenge. Researchers have come up with several ideas to address this crucial problem. However, current models still face challenges in balancing scalability and structural consistency. This study proposes a feature imputation method, called FILLER, that deliberately searches the two-dimensional latent space produced by a generative model and fills the missing values with appropriate entries. The generative model is trained on fully observed data to generate samples from the latent space, and FILLER uses this trained model to impute the values missing in the corrupted test samples. In this study, G-NeuroDAVIS serves the purpose of the generative model. This work also presents a mathematical proof on the convergence of the iterative search. Finally, FILLER has been evaluated on several image datasets under random and structured missingness patterns with varying levels of imputation complexities. In order to justify the efficacy of FILLER, it has been compared against existing state-of-the-art solution strategies in terms of RMSE, PSNR, and SSIM. In addition, Wilcoxon signed-rank test has been carried out to validate statistical significance. Moreover, downstream analyses (classification and clustering) have also established the quality of imputation in terms of standard metrics.

AI-generated content (AIGC) has made significant progress, with 2D generative models becoming ready-to-use tools for the digital fashion industry. However, 3D garment generation remains in its nascent stage, where in the realm of fashion, the semantic information of diverse design elements exhibits intricate coupling relationships in 3D representations, posing substantial challenges for generating diverse 3D garments. In this work, to handle the above problem, We introduce Fashion-3DLR, a novel 3D garment generation framework that utilizes diverse design elements to create high-quality, versatile 3D garment assets. Specifically, to bridge the semantic gaps between different fashion elements, we propose a Garment Feature Fusion Diffusion Transformer (GFF-DiT) module to integrate 2D fashion design elements, e.g., sketch and texture, into latent space. Within the latent space, we then employ a rectified flow transformer to generate geometry latents, which can be decoded into various 3D garment representations, including 3D Gaussians and meshes. Furthermore, we integrate Fashion-3DLR into downstream tasks, achieving the 3D Gaussian Splatting (3DGS)-driven cloth physical simulation and mesh-based virtual try-on. Experimental results indicate that Fashion-3DLR surpass the previous state-of-the-art methods, which verify that the proposed work can generate well-structured, non-watertight garments capable of physical simulation and virtual try-on, underscoring its potential as a versatile 3D garment design tool.

Multi-head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value pairs through a shared low-rank bottleneck (cKV), achieving 81% KV-cache reduction during inference. Despite its adoption in massive production models, no prior work has studied what information this bottleneck preserves or discards, nor how it reshapes internal transformer circuits. We present the first comprehensive mechanistic interpretability study of MLA, training a 114M-parameter transformer (pretrained on a web/code/math mixture, fine-tuned on TinyStories) and analyzing its representations through SVD, attention head taxonomy, linear probing, and a disruption-attribution analysis. Our key findings are: (1) the cKV bottleneck learns a pure content representation, preserving entity identity (98% retention) while discarding positional information, validating MLA's separation of content from position via RoPE; (2) induction heads co-locate at a single layer (Layer 12), unlike their distributed formation in standard MHA; (3) a single "semantic hub" layer (Layer 15) simultaneously exhibits the highest SVD effective rank and strongest disruption-attribution score; and (4) the bottleneck is globally over-provisioned, using only 46% of its capacity on average. These findings suggest MLA does not merely compress attention passively, but reshapes how the model organizes content, position, and circuit structure. We view this as an initial data point and detail scope limitations in Section 5.

Multimodal embedding spaces in models like CLIP enable powerful capabilities such as semantic similarity retrieval and cross-modal zero-shot classification. These embeddings compress high-level semantics into a single vector, which comes at the cost of primarily expressing a dominant semantics like main object while suppressing other important attributes such as camera angle or color tone. We propose a text-conditioned transformation of visual embeddings that makes such attributes explicitly accessible. Given a natural language description of an attribute category (e.g., "color" or "art style"), a network generates an affine transformation that emphasizes the specified attribute. Conditioning on text enables it to learn many attributes simultaneously, accessing them at inference time through an intuitive interface. The network is trained to align transformed embeddings with the frozen latent space, enabling retrieval using existing large-scale embeddings without any re-encoding. When applied to a full set, the same mechanism transforms the latent space for attribute disentanglement tasks such as multi-clustering. By operating directly in latent space, our method provides a unified and efficient framework for controlling embedding spaces, demonstrating state-of-the-art performance across both attribute-based retrieval and multi-attribute organization tasks with near-zero inference cost. Project page: https://joefioresi718.github.io/ControlEmbed_webpage/

Joint Embedding Predictive Architectures (JEPA) have recently emerged as a paradigm for learning world models by predicting latent representations, offering a promising direction for self-supervised learning. While initial attempts have applied JEPA to the music domain, it remains unclear how such frameworks can naturally support the formation of a world model for music. In this work, we propose to learn a world model of piano sound using JEPA by framing music as an action-conditioned system: the audio is treated as the state, and the pianoroll as the instrument action. Given a current audio state and an action, the model predicts the resulting future audio state, mirroring how humans learn musical sound through interaction. The model is trained in a fully offline setting using paired audio-pianoroll data, without environment interaction. Experiments show that the learned model captures the relationships between musical actions and their resulting sound. The resulting representations support downstream tasks, including beat tracking, composer identification, and key estimation, and enable piano transcription via planning, by searching for actions that best explain a target sound.

Neural audio codecs are challenging transformations for audio watermarking because they re-encode, quantize, and resynthesize speech. This paper investigates continuous latent-space watermarking for codec robustness. Instead of adding a watermark only to the waveform or spectrogram, we embed a 32-bit message into the continuous latent representation of a codec-like speech autoencoder. The pipeline uses a SEANet-style encoder-decoder, a Conformer-based message embedder, RVQ-guided latent decomposition, and a latent-domain detector trained under signal-processing and neural-codec transformations. Rather than proposing a final universal watermarking baseline, we characterize the trade-offs that appear when the watermark carrier is moved before neural decoding. On 48 kHz speech, EnCodec-aware training improves EnCodec-24k bit accuracy from 78.8% to 95.6% and 97.1%, while PESQ decreases from 3.727 to 3.514 and 3.427.

We propose FA-LAM, a Focus-Aware Large Avatar Model for one-shot animatable Gaussian head creation, while simultaneously enabling static 3D and dynamic 4D full-head recovery. The core of our method lies in a thorough analysis of the attention mechanisms and the entangled reconstruction and animation training pipeline adopted by prior state-of-the-art approaches. Our analysis identifies two main factors that compromise the quality of 3D full-head generation: (1) incorrect and noisy attention activations, and (2) conflicts between the tasks of reconstruction and animation. To address the first issue, we introduce a symmetric and semantic attention regularization strategy that leverages the inherent semantics and structural symmetry of human heads. To disentangle the objectives of reconstruction and animation, we develop a novel dual-phase training pipeline that separates the model's capabilities for large-view hallucination and animation into distinct modules. Moreover, we enhance our model to support multi-view and streaming 4D reconstruction in an efficient and memory-friendly manner through a core autoregressive modification with tailored visibility-aware token fusion. Collectively, these innovations enable FA-LAM to reconstruct animatable Gaussian full heads with superior quality, particularly in fine facial regions and large viewing angles.

A recent line of work measures causal emergence in reinforcement learning agents through Integrated Information Decomposition, reporting that $Φ_r$ grows with training and tracks reward improvement. For active inference, this raises the question of how reward-free predictive organization relates to such information-theoretic signatures. I test this within an active inference agent whose architecture separates a fast perception latent $z$ from a slow global latent $g$, where $g$ is driven by prediction error and structurally decoupled from policy gradients. In a reward-free environmental regime-switching protocol, $Φ_r$ concentrates in $g$; its aggregate magnitude is largely architectural and decreases with training. The substantive effect of learning becomes legible only at the atom-compositional level: decoupling flips sign from negative to positive and becomes regime-invariant under environmental change, while downward causation carries the regime-dependent adjustment. These results identify $g$ as the architectural locus of $Φ_r$-relevant temporal organization in an active inference agent, and argue against reading scalar $Φ_r$ as a direct index of learned integration.

Modeling continuous object deformation is important for many computer vision and robotics tasks, such as manipulation and simulation. Existing approaches rely on learning-based methods or physics simulators to model shape deformations. However, these approaches either use discrete time steps or are too computationally intensive for real-time applications. We present ODeform, a novel extension of Neural Ordinary Differential Equations to continuous 4D dynamics of deformable objects in 3D space. Our method transforms 3D point clouds and physical conditions (like material properties) into a unified latent space. By solving the resulting ordinary differential equations over time, we model deformations as continuous flows within this learned embedding, eliminating the need for discrete time steps while maintaining computational efficiency. We evaluate our approach on unseen physical parameter configurations, showing improved motion prediction accuracy over baseline methods. Our experiments further demonstrate a successful transfer to real 3D captured objects with novel shapes, along with effective interpolation and extrapolation of the learned dynamics. Our code and data will be made publicly available.

We propose the effective rank (ERank) of the channel covariance of an image's deep feature map as a per-sample, label-free measure of visual richness, computed from a single forward pass through a frozen pretrained encoder. ERank counts how many decorrelated channel directions an image activates, and we characterize its properties, including its behavior under noise. Empirically, ERank orders images from plain to visually rich, correlates with codec bitrate, sharpness, and edge density, and correlates with human complexity annotations on IC9600 with $r = 0.72$. As a data-selection criterion, removing low-ERank samples improves super-resolution and removing high-ERank samples improves OCR, in both pretraining and finetuning, while selection does not help classification, segmentation, or denoising. ERank is thus a cheap richness signal, useful exactly when task difficulty is governed by input richness.

Most existing video compression algorithms follow a paradigm of transformation and quantization, optimizing the trade-off between distortion and bitrate. However, extremely low-bitrate compression remains an underexplored frontier where perceptual quality optimization under severely constrained coding resources has not been adequately addressed. In this paper, we propose a unified generative framework that leverages pre-trained Diffusion Transformer (DiT) priors to achieve high perceptual quality at extremely low bitrates. We first introduce a flexible Group-of-Latents (GoL) strategy within the latent space of a causal tokenizer, explicitly partitioning the latent stream into intra $I$-latents and inter $P$-latents. The Deep Compression Module (I-DCM) then encodes key $I$-latents to preserve perceptual anchors with minimal overhead. Building upon these anchors, the DiT-based Unified Latent Denoising Module (U-LDM) refines intra-frame textures and synthesizes $P$-latents from noise, reconstructing temporal dynamics at zero additional bitrate cost. Extensive experiments demonstrate that our method uniquely operates in the extreme-low-bitrate regime (e.g., (<0.005) bpp), achieving state-of-the-art perceptual fidelity with rich spatial details and robust temporal consistency. The code will be made publicly available.

Generative models trained on a source domain often produce samples that are poorly aligned with shifted target domains, limiting their effectiveness for target-domain data augmentation. Although target-specific adaptation can reduce this mismatch, it typically requires additional optimization and domain-specific parameters. We propose a Similarity-based Generative Network (SGN), a reusable framework that is trained once on labeled source data and applied to new target domains without parameter updates. SGN learns a latent space structured by label-induced pairwise similarities while preserving reconstructive information through an encoder-decoder architecture. At generation time, a small labeled representative set from the target domain is encoded and combined in the learned latent space, allowing the generated samples to inherit target-specific characteristics while maintaining class consistency. We further analyze the realizability and dimensionality requirements of the proposed similarity structure. Experiments on image and tabular datasets demonstrate the effectiveness of SGN for target-guided data augmentation under source-to-target distribution shifts.

Establishing interpretable decision-making processes in long-horizon robotic manipulation is critical for enabling reliable human oversight and intervention. However, existing approaches to robotic manipulation largely treat skill selection as opaque mappings from observations to actions, offering limited transparency into how decisions are formed. In this work, we propose ConceptTree, a framework that reframes high-level manipulation skill selection as reasoning over human-interpretable concepts, representing high-level policies as a sequence of concept-level predicates over visual observations. Rather than relying on implicit latent representations, our method learns a normalized concept space grounded in visual inputs, over which a decision tree is trained to predict high-level skills. This formulation yields a transparent decision process that is both traceable and intervenable, enabling direct inspection and modification of policy behavior. We evaluate our approach on a set of real-world robotic manipulation tasks with increasing complexity. Experimental results show that ConceptTree consistently outperforms existing concept-based baselines, particularly in complex, long-horizon scenarios. Furthermore, we provide qualitative case studies showing that our model supports fine-grained intervention by modifying individual concepts, enabling targeted correction of decision errors without retraining.

Detecting high-level semantic concepts like negation across modalities remains a challenge for current multimodal systems. We analyze this as a fundamental representation learning problem, providing the first evidence that negation does not form a linearly or non-linearly separable class in the latent spaces of standard vision-language models (VLMs). We demonstrate that pretrained embeddings primarily encode modality-specific features, lacking a generalizable negation signal. To overcome this, we propose a novel cross-modal attention architecture that explicitly models inter-modal dependencies, achieving performance gains of up to +7.03% F1 over unimodal baselines. Our analysis reveals a key asymmetry: while textual negation often appears independently, visual negation is semantically dependent on linguistic context, a finding validated through our statistical analysis of 3,222 political video-text pairs automatically annotated via \textsc{Qwen2.5-VL}. By combining this analysis with self-supervised video representations (JEPA2), we advance the modeling of temporal negation. This work provides new methods and insights for learning robust, semantically-aligned representations in multimodal systems.

Continuous Chain-of-Thought methods replace verbose reasoning traces with a short sequence of dense latent representations. Earlier continuous CoT methods indirectly supervise the latent representations such that its final state match that of verbose reasoning traces, requiring autoregressive, slow generation during training. We introduce C-MTP, a simpler, faster direct supervision approach that models each latent as an average of the embeddings in the CoT traces to be compressed. Our approach outperforms a prior direct supervision method that approximates the distribution of compressed tokens, and performs competitively to slower indirect supervision approaches in existing evaluation setup with simplified CoT traces (less than 100 tokens). Lastly, we extend the evaluation of Continuous CoT methods to complex tasks with longer reasoning traces ($\ge$ few hundreds reasoning tokens). We find both direct and indirect supervision training methods perform poorly (roughly 65% performance drop) in this setting, revealing the limitations of current continuous CoT methods. The code and checkpoints are released at https://github.com/Varun221/cmtp_research

As Large Language Models (LLMs) grow more capable across diverse tasks, their (in)ability to generalize remains difficult to quantify and poorly understood beyond limited domains. In particular, LLMs are known to struggle generalizing multilingually, to languages outside of English, and that are poorly attested in their training data. To understand why this may be, and what enables some models to perform better than others, we turn to a long history of work across the cognitive sciences, arguing that successful generalization derives from appropriate representations in similarity space. We look at how well LLMs' representations capture the hierarchical similarity structure between distinct languages. Strikingly, we show LLMs' latent representations largely recover the hierarchical structure of the Indo-European language family tree -- grouping languages that are members of the same subfamily closely together in representation space. Furthermore, we show that the degree to which models reflect the similarity structure of languages correlates with their performance on XNLI, a multilingual natural language inference benchmark. This extends classic work on similarity-driven generalization at scale, showing how models that represent similar languages similarly generalize better from one language to another.

Imitation Learning aims to learn skills from extensive observations and demonstrations for robots, so it suffers from data scarcity and environment generalization. The existing methods predominantly focus on imitation from in-domain tasks and consequently struggle with generalization to unseen tasks. To bridge this generalization gap, we propose the \textbf{D}ynamics-\textbf{A}ware \textbf{M}eta-\textbf{I}mitation (DAMI) framework. By integrating meta-learning to construct a shared skill space, DAMI equips agents for rapid adaptation to novel tasks. We introduce the Visual-Motor Trajectory (VMT) module to capture complex spatio-temporal dynamics within the task latent space. Furthermore, we propose the Unpaired Unified Task (U2T) block to fuse unstructured multimodal observations. To coordinate these representations, we integrate a Task-Conditioned Feature Modulation (TCFM) mechanism customized for modulating low-level 3D features. By capturing intrinsic dynamics from a random complete reference demonstration, our framework learns the underlying task logic rather than memorizing static cues, ensuring effective generalization. Extensive experiments in both simulation and real-world settings demonstrate that our approach outperforms state-of-the-art baselines regarding direct inference on seen tasks and adaptation to unseen tasks via few-shot fine-tuning.

Video generative models commonly rely on latent spaces learned by 3D Variational Autoencoders (3D-VAEs). However, conventional 3D-VAEs are mainly optimized for pixel-level reconstruction, which can limit the semantic and spatio-temporal structure captured by their latents. Meanwhile, Video Foundation Models (VFMs) such as V-JEPA 2 and VideoMAEv2 show strong video understanding capabilities, yet whether their frozen representations can be transformed into compact, reconstruction-capable, and generation-friendly video latents remains largely unexplored. We answer this question with VideoRAE, a representation autoencoder that leverages multi-scale hierarchical features from a frozen video foundation encoder and compresses them with a lightweight 1D self-attention projector. VideoRAE supports both continuous latents for Diffusion Transformers and discrete tokens for autoregressive models via multi-codebook high-dimensional quantization. During decoding, a local-and-global representation alignment objective with the frozen VFM teacher improves semantic preservation and enables training without KL regularization. Experiments show that VideoRAE achieves strong reconstruction in both continuous and discrete regimes. On UCF-101, it obtains state-of-the-art class-to-video gFVDs of 40 and 93 with AR and DiT generators, respectively, while converging approximately 5x faster than competing autoencoder baselines. In a controlled 2B-scale text-to-video study, replacing LTX-VAE with VideoRAE leads to faster convergence under comparable settings. These results validate frozen VFM representations as versatile and generation-friendly video latents. The model and code will be released on https://zhxie0117.github.io/VideoRAE.

Video diffusion transformers (vDiTs) generate high quality video but introduce extremely high compute cost due to the long diffusion timesteps and self attention computation. As diffusion timesteps are reduced, the computation cost of self attention becomes the dominant bottleneck. Existing acceleration approaches largely inherit sparse attention techniques from large language models, which fail to consider the unique spatiotemporal correlation of video data. This paper presents Kaleido, an algorithm hardware codesign that accelerates all operations in vDiTs by exploiting channel-wise spatiotemporal correlations in latent space. Based on this insight, we propose a lightweight channelwise reuse algorithm that skips redundant computations by reusing partial results while preserving higher generative quality than prior methods (>17 dB). To efficiently support this algorithm, we design a systolic array like accelerator with reconfigurable processing elements and a lightweight data dispatcher to mitigate irregular sparsity and data access patterns introduced by our reuse algorithm. Evaluations across three mainstream vDiT models show that Kaleido achieves up to 5.9x speedup and 16.0x energy savings over state of the art accelerators.

Life-limiting congenital anomalies require accurate prenatal diagnosis for appropriate clinical decision-making. Prenatal ultrasound (US) examinations involve multiple anatomical planes, and diagnosis depends on identifying anatomical planes and selecting diagnostically relevant planes for each anomaly. Existing automated methods either rely on plane-level annotations or aggregate heterogeneous images without explicitly modeling these diagnostic capabilities. We propose AnomExpert, a prototype-driven framework for prenatal US anomaly diagnosis using only case-level supervision. AnomExpert introduces learnable plane prototypes to organize unordered images into latent representations corresponding to anatomical planes without requiring plane annotations. A disease-aware sparse selection mechanism further selects diagnostically relevant planes for each anomaly. Experiments on a multi-center dataset of 3,654 cases show that AnomExpert consistently outperforms nine representative multi-instance learning methods. Using a ViT-small backbone, it achieves 86.9% accuracy and 84.2% F1-score while maintaining parameter efficiency. These findings indicate that modeling anatomical plane identification and disease-specific plane selection improves weakly supervised multi-plane prenatal US anomaly classification. The code is available at https://github.com/TIanCat/AnomExpert.

Multimodal semantic segmentation (MSS) is essential for robust perception in complex environments, yet its potential remains largely untapped because of the prohibitive cost of human annotations. While unsupervised semantic segmentation (USS) has achieved strong results on a single RGB modality, its naive extension to multimodal data is often hindered by fusion degradation. This occurs because, without explicit supervision, existing frameworks struggle to reconcile the heterogeneous structural patterns captured by different sensors and therefore fail to effectively exploit their complementary information. In this paper, we make the first attempt to address the novel problem of Unsupervised Multimodal Semantic Segmentation (UMSS), aiming to effectively exploit complementary sensor information in a fully label free setting. To this end, we propose UniM2 (Unified Multimodal), a novel framework built on DINOv3 that transforms conventional fusion methods into consistent performance gains. Our key idea is to learn a unified latent space driven by Cross Modal Correspondence Synergy (CMCS) to extract intrinsic shared semantic cues, bypassing the need for label guided adaptive fusion. To mitigate inherent intermodal conflicts, we introduce a Cross Modal Harmonizer (CMH) that designates RGB as a stable reference, effectively suppressing inconsistent relational supervision while guiding the model to exploit complementary structural features. Extensive experimental results on NYU Depth v2 and MFNet show that UniM2 improves mIoU by 6.4% and 9.8%, respectively, demonstrating clear advantages over existing frameworks for UMSS.

Graph Foundation Models (GFMs) have emerged as a promising paradigm for learning transferable representations across diverse graph domains. Recent advancements in GFMs have been largely dominated by two paradigms: Graph Neural Network and Large Language Model (LLM) based methods. However, these methods often face a fundamental dilemma between training with limited data and a heavy reliance on textual attributes. Tabular foundation models (TFMs) offer a potential alternative, as node features and representations can be naturally organized in a tabular form. However, how to enable TFMs to effectively capture structural information of graphs remains largely unexplored. The key challenge is to learn a graph-to-table alignment mechanism that enables graph structural understanding for TFMs. To address this, we propose GTAlign, a surprisingly simple yet effective Graph-to-Table Alignment framework for text-free Graph Foundation Model. Specifically, we first pretrain a graph encoder that maps diverse graphs into a unified latent space to capture domain-agnostic graph representations. To further bridge the gap between graph topology and the tabular representation space, we propose community-guided continual pre-training, where pseudo-labels derived from graph community are used to construct few-shot prediction episodes. Lastly, we adapt the graph encoder for an unseen target domain and perform in-context inference. Extensive experiments on five benchmark datasets demonstrate that GTAlign significantly outperforms state-of-the-art baselines on both node and graph classification, offering a simple, effective, and text-free GFM model. Code will be released upon acceptance.

AI music generation has rapidly advanced alongside commercial platforms, raising the need for reliable watermarking for provenance and attribution. However, existing audio watermarking research has largely focused on speech, and applying speech-oriented methods to music is challenging due to music's complex structure and rich acoustic texture. Most existing methods are post-hoc, adding imperceptible perturbations after generation rather than embedding watermarks as part of the content. This makes them fragile under transformations and especially vulnerable to neural codec re-synthesis, which can discard imperceptible residual signals. Moreover, since generation and watermarking are decoupled, the watermarking step can be bypassed or omitted, weakening provenance guarantees. To address these issues, we propose MusicMark, which, to the best of our knowledge, is the first generative watermarking framework for music. Specifically, MusicMark embeds watermark messages into the semantic latent space during generation, incorporating the watermark as part of the musical content and ensuring robustness against diverse attacks, particularly neural codec re-synthesis. To this end, we introduce a watermark adapter into a diffusion-based generation model to embed watermark messages across denoising steps. The adapter and detector are trained with a joint objective that preserves fidelity by constraining watermarked latents close to their unwatermarked reference latents, while improving robustness through attack augmentations. Experiments demonstrate that MusicMark substantially outperforms post-hoc baselines across diverse attacks including neural codec re-synthesis, while maintaining comparable generation quality. We further introduce a cover-song attack, converting the singing voice while preserving musical content, and show that MusicMark remains more robust than post-hoc methods.

Large language models (LLMs) often produce homogeneous outputs, raising concerns that AI coding assistants may lead to convergence in the software artifacts that developers create. Whether this occurs in practice is unclear because developers interactively prompt, evaluate, modify, and reject model outputs, and because outputs vary with prompt and repository context. I examine code homogenization using Kaggle contest submissions from 2019 to mid-2026. I first document widespread convergence toward the random seed value 42, consistent with LLMs reinforcing a longstanding convention in programming culture. I then study homogenization more broadly, at two levels of aggregation and abstraction. At the submission level, I measure the average pairwise similarity of submissions within contests. At the contest level, I measure the conceptual span of submitted code, motivating distinct measures for each: TF-IDF representations, which capture surface syntax, and Voyage 3 code embeddings, which capture code intent and semantics. The results demonstrate substantial syntactic homogenization at both the individual and collective levels: individual submissions have become more alike in literal syntax and code structure, while the latent dimensionality of syntactic variation has narrowed. In contrast, I find little evidence of semantic homogenization, individually and collectively. Average semantic distance remains essentially flat, and the contest-level latent dimensional span of semantic approaches remains stable, with evidence suggesting it has even expanded modestly. These findings suggest that AI coding assistants are certainly standardizing implementation details, yet they have not yet produced evidence of homogenization in the approaches and problem-solving strategies coders employ.

Flow matching over carefully designed latent representations has recently emerged as a powerful paradigm for topology-aware mesh generation. Existing approaches, however, model vertices and connectivity jointly in a joint latent space, entangling continuous vertex geometry with discrete combinatorial structure; this complicates flow learning and manifests as drifting vertices and broken surfaces. We present LATO.2, a factorized flow matching framework that decomposes mesh generation into a vertex flow followed by a connectivity flow conditioned on the realized vertices, with both stages anchored to a shared coarse voxel scaffold. Dedicated VAEs underpin the two stages, recovering vertices at sub-voxel precision and embedding discrete connectivity into a continuous latent space. We demonstrate two advantages unique to this factorization: (i) part-wise generation, in which the scaffold is partitioned and each part synthesized at full latent capacity, yielding substantially higher-resolution meshes than a monolithic latent permits; and (ii) topology-adaptive editing, in which manipulating first-stage vertices induces the corresponding connectivity without re-optimization. Experiments show that LATO.2 surpasses state-of-the-art topology-aware mesh generators in geometric fidelity and connectivity quality.

We introduce DenseAR, a new generative paradigm that reformulates autoregressive image generation as coarse-to-fine next-dense-stride prediction using a compact single-scale tokenizer. Our key insight is that traversing a single-scale latent grid with progressively denser strides naturally captures the transition from global structure to fine detail. This addresses two limitations of existing autoregressive models at once: the slow inference of raster-order autoregression, which DenseAR avoids by predicting multiple tokens in parallel, and the heavy cost of multi-scale approaches, which need long, multi-resolution token sequences to achieve coarse-to-fine prediction. Building on our efficient framework and the flexibility of autoregressive modeling, we further extend DenseAR to a unified model that handles multiple modalities and imaging tasks within a single backbone. We validate DenseAR on both medical and natural images. On multi-contrast brain MRI, a single DenseAR model unifies cross-modal translation, modality-conditioned generation, and tumor segmentation, while remaining competitive with task-specific methods. On ImageNet, DenseAR improves class-conditional generation quality (FID and IS) over both a single-grid baseline without stride ordering and a multi-scale tokenizer-based baseline.

Dataset condensation for action segmentation synthesizes compact, informative representations of long, untrimmed video datasets. The existing approach relies on Variational Autoencoders and an iterative latent optimization; it is computationally expensive and suffers from over-smoothed reconstructions and rigid temporal constraints. This paper proposes to shift the condensation paradigm from optimization-based inversion to deterministic latent mapping. By leveraging Denoising Diffusion Implicit Models, we represent action segments as continuous trajectories anchored by sparse latent points in the noise manifold. To maximize representational efficiency, we introduce an adaptive allocation mechanism that dynamically redistributes the anchoring budget based on segment-wise reconstruction difficulty. Extensive experiments demonstrate that our framework significantly outperforms state-of-the-art methods in segmentation performance across common datasets. Notably, our approach achieves performance parity with real data training while maintaining a condensation ratio of 2.4% on Breakfast dataset.

Phone segmentation and recognition are inherently related tasks, yet modern approaches typically model them separately. We argue that phonetic structure is already latent in the representations of self-supervised speech models (S3Ms), and one only needs to steer them to solve both tasks. We leverage S3M-based Phonological Activation Mapping (SPAM), which maps each S3M representation frame to a vector of phonological feature activations, such as voicing and nasality. On top of SPAM, we introduce two simple but effective lightweight, gradient-descent-free prediction heads: a recognition head and a segmentation head. Our method requires less than a minute of phonetic transcriptions, and generalizes to unseen phones during training. Across a diverse range of datasets, our approach attains strong segmentation and recognition performance.

Pretrained generative robot policies based on flow matching and diffusion have achieved impressive results across a wide range of manipulation tasks. Yet real-world deployments routinely expose failure modes outside the pretraining distribution. Closing these gaps typically requires large-scale data collection or online reinforcement learning on physical hardware, which is impractical for rapid and safe adaptation. We present FlowDAgger, a sample- and compute-efficient method for adapting frozen generative robot policies from human interventions in latent space. Our key idea is action inversion: each human expert action is mapped to the noise that would have produced it under the frozen base policy, using reverse-time integration followed by local refinement. The resulting inverted noise provides supervision for a lightweight latent policy that steers the base model at deployment time, enabling rapid skill acquisition while preserving its behavioral priors. We evaluate FlowDAgger in simulation and on real-world bimanual and single-arm manipulation, adapting both action-head VLAs and world-action models from a handful of interventions. FlowDAgger outperforms supervised fine-tuning and latent-space RL baselines and preserves pretrained skills on held-out tasks, offering a practical path for adapting robot foundation models in the real world. Website: https://microsoft.github.io/FlowDAgger

End-to-end neural audio models achieve high-fidelity compression and generation. We might read that performance as evidence they directly represent interpretable features such as pitch and timbre, but a model can produce plausible outputs without doing so. A model may encode these features in any reachable basis, but regardless of which, the features are well described as compositions of time-frequency-localized primitives. Whether state-of-the-art encoders preserve access to these primitives, and thus to compositions of them, remains unclear. Through theoretical analysis and controlled experiments, we show that several state-of-the-art strided convolutional encoders impose two structural bottlenecks, both predictable from architecture and signal structure, on access to these primitives: (1) they collapse primitives into alias equivalence classes, establishing a bound on representational capacity, and (2) they limit the frequency resolution available to learned filters, restricting separability. For well structured data, we find collapse rates of 31-35% and filter bandwidths 10-35x above the theoretical resolution bound, confirming that both bottlenecks arise under realistic signal conditions. We then introduce Gabor Latent Refactorization (GLRF), a lightweight post-hoc intervention that re-expresses encoder latents in a frequency-localized basis, reducing filter bandwidths from 10-35x to 1.5-3x of the theoretical resolution bound while preserving reconstruction fidelity and improving control over attributes like pitch. These results show that the encoders in question predictably degrade access to frequency-localized primitives, entangling the features that depend on them, and that a lightweight, retraining-free intervention can recover much of that access, improving steerability and interpretability.

Scaling Large Language Models (LLMs) has been driven mainly by enlarging the Transformer backbone, but for an already-strong model this requires another round of costly pretraining. We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed. Depth-recurrent (looped) Transformers pursue this goal but are hard to scale, because looped computation does not fit naturally with the pipeline parallelism used to train the largest models. We add computation along the sequence-length dimension, where the extra computation is simply a longer input and stays compatible with standard large-model training. We propose Hidden Decoding, a sequence-length scaling method applied during continued pretraining (CPT). It expands each token into n streams with independent embedding tables and keeps the intermediate streams' key-value cache as context, so each token performs more internal computation without adding or widening Transformer layers. To keep this affordable at scale, we introduce Stream-Factorized Attention, in which most layers attend only within each stream and only a few layers mix across streams, reducing the attention cost from quadratic to roughly linear in n. Experiments support two scaling results. At frontier scale, we train WeLM-HD4-80B and WeLM-HD4-617B at n=4 and improve their matched non-HD baselines, making Hidden Decoding the first demonstrated sequence-length scaling method at the 100B+ MoE scale. Across expansion factors, the gains grow as n increases, showing that sequence-length expansion is a practical fixed-backbone scaling path for frontier-scale LLMs.

This study investigates the presence and propagation of bias within Neural Networks through a comprehensive multi-level analysis spanning the learned latent space, layer activations, and the network's parameters. Based on this taxonomy, we propose three bias detection approaches: 1) SpaceBias (new method), which characterizes the latent space prior to the final classification layer using neighbor-probability distributions and quantifies bias with the two-sample Kolmogorov-Smirnov test on the per-group distributions. 2) ActivationBias (extension of the existing method InsideBias), which analyzes the activations of neural network filters and quantifies bias via a Mann-Whitney U test, based on the observed fact that underrepresented groups exhibit lower activation levels in the final convolutional layers. 3) WeightBias (extension of the existing method IFBiD), which uses a secondary neural network trained to identify biased patterns directly in the parameters of task-specific models. Unlike conventional methods, which assess neural network outcomes and treat the model as a black box, our proposed techniques provide insight into how biases manifest within the network architecture itself at different levels, offering a more nuanced and detailed understanding. Experiments are conducted on two complementary applications: gender classification in the DiveFace dataset (72,000 face images) and digit classification on a colored-MNIST benchmark with controlled bias severity. In total, more than 127,000 models with varying degrees and types of bias were trained and evaluated. The severity sweep shows that the internal disparity, and with it the detection performance, decreases smoothly as the training distribution approaches balance. The results highlight the importance of methods that provide deeper insight into the behavior of AI models.

Lattice field theory is the workhorse of non-perturbative physics, used to simulate phenomena from the strong nuclear force to critical phenomena in materials. Its Boltzmann distributions are parametrized analytically by coupling constants, but these bare parameters are weak predictors of observables -- extracting physics typically requires extensive simulation. While normalizing flows have emerged as effective samplers at fixed couplings, it remains difficult to interpret what these networks have learned. This raises a natural question: can the physics be read off directly from the flow network parameters themselves, and can those parameters be generated for unseen theories? We propose lattice field theory as a testbed for neural network interpretability: because the target physics is qualitatively well-understood and smoothly varying, it provides ideal synthetic data with known ground truth. To this end, we introduce JEPAWG, a Joint-Embedding Predictive Architecture-based Weight Generator that maps couplings directly to flow weights via a learned latent space. On a scalar theory at lattices of size $6^2$ to $11^2$, the JEPAWG latent space recovers the correct intrinsic dimension of the underlying manifold, locates the phase transition, and encodes a finite-size shift aligned with the 2D Ising exponent $ν\approx 1$, allowing us to uncover physical structure by studying the network weights alone. This suggests the fascinating idea of treating the network weights as a new type of physical observable. As a generator, JEPAWG also interpolates and extrapolates to unseen couplings effectively and remains robust to weight-space discontinuities introduced by multi-seed training data, outperforming PCA, AE, and VAE baselines.

Unified 3D foundation models aspire to generate 3D assets and reason about them in language within a single backbone, but their text-3D interaction remains largely implicit. Existing methods concatenate text and 3D tokens into a flat sequence and rely on self-attention, collapsing coarse structural cues and fine geometric details into one undifferentiated representation. We introduce ELSA3D, a unified 3D model that addresses this with elastic semantic anchoring, structuring language and geometric reasoning jointly along matched abstraction scales. ELSA3D represents geometry with a scale-aware octree tokenizer and introduces Anchor Tokens, sparse cross-modal units that select semantic cues, route them to the most relevant 3D scale, retrieve scale-specific geometric evidence, and write the fused signal back into the unified representation, keeping interaction sparse yet precise. A lightweight per-block router makes both computation and reasoning elastic, choosing which text tokens instantiate anchors at which geometric scale so that cross-modal capacity concentrates where alignment is most needed. ELSA3D achieves state-of-the-art performance across image-to-3D generation, text-to-3D generation, and 3D captioning, outperforming the strongest unified baseline while roughly halving FLOPs and inference latency relative to the non-elastic version of the same model.

Video-to-audio (V2A) generation aims to synthesize realistic audio that is both semantically consistent with and temporally synchronized to a silent video. Despite recent progress, many methods still rely on multi-stage training, resulting in high computational costs and long runtimes, or transform visual input into text to leverage pretrained text-to-audio models, sacrificing fine-grained temporal cues. To overcome these limitations, we propose Flowley, an end-to-end, single-stage training architecture that produces soundtracks by combining visual features with textual prompts. Crucially, we introduce Progressive Soft-masked Cross-Attention, which embeds audio-visual synchronization directly within its attention mechanism, adding zero additional computational cost compared to standard attention layers. We further observe that existing V2A benchmarks lack sound-oriented descriptive captions, which can potentially degrade the quality of the synthesized audio. To remedy this, we propose SoundCap, a plug-and-play pipeline for creating detailed, sound-aware captions that guide the model. Remarkably, without integrating any pretrained audio-visual alignment modules, Flowley achieves state-of-the-art performance on VGGSound across multiple metrics. Moreover, by incorporating SoundCap, we further exceed the performance of the strongest existing close-sourced methods in terms of audio quality in the zero-shot setting.

Recent diffusion-based generative models have shown impressive performance in image generation and editing. However, due to memory limitations and the high cost of collecting high-resolution training images, existing methods are typically restricted to inputs with linear resolutions below 1K. In contrast, photos captured by modern mobile devices often reach linear resolutions up to 8K, revealing a significant gap between current capabilities and real-world demands. Simply upscaling low-resolution edited results often results in visually enlarged but blurry images that lack fine details. This paper introduces UltraDiffEdit, a novel, tuning-free image editing framework that extends off-the-shelf latent diffusion models (LDMs) to ultrahigh resolutions. UltraDiffEdit employs a multi-scale progressive editing strategy, iteratively blending high-resolution edited content with unedited areas in a coarse-to-fine manner. We employ multi-patch encoding to preserve both edited and unedited visual details within the latent space. To mitigate editing artifacts, our global-local consistency denoising technique consistently integrates edited and unedited latent features, ensuring smooth transition at editing boundaries from the latent representation to the final image. We also introduce a patch-based hybrid sampling approach that captures local, intermediate, and global features, ensuring semantic coherence and enhancing fine detail during denoising. We conduct extensive experiments demonstrating UltraDiffEdit's superior editing quality and flexibility: it can handle image resolutions up to 8K using only a single NVIDIA GeForce RTX 3090 GPU. The source code is publicly available at https://github.com/LonglongaaaGo/UltraDiffEdit.

Synthesizing physically plausible human-scene interactions (HSI) remains a critical challenge in computer vision and the development of human avatars. Although recent generative models enable diverse motion synthesis, they suffer from an inductive bias referred to as semantic-geometric entanglement. Because spatial constraints often strongly correlate with specific actions in training data, monolithic models will learn the shortcut bias, aggressively overriding the semantic intent when faced with strict geometric cues. Furthermore, this entanglement exacerbates physical hallucinations, such as body-scene penetrations. To address these limitations, we propose DeSeG, a hierarchical framework that explicitly decouples semantic intent from geometric constraints. First, we introduce a Residual Semantic Planner that encodes textual instructions and canonicalized goal voxels into a compact latent space, enabling fine-grained semantic control independent of spatial trajectories. Second, we propose a physics regularized diffusion executor that incorporates differentiable repulsive potential fields directly into the diffusion objective, enforcing collision-aware motion generation. Extensive experiments on the Lingo dataset demonstrate that DeSeG achieves state-of-the-art performance, reducing mean scene penetration by 47% and improving semantic alignment by 29% over the SOTA baselines.

Principal Component Analysis or PCA-like properties (orthogonality, variance ranking) are seldom realized in deep autoencoder architectures. In this work, we present ODIN (Orthogonal Dendritic Intrinsic Network), a novel autoencoder architecture that recovers PCA-like latent structure in a fully non-linear regime. By incorporating a set of geometric constraints directly into the training objective, ODIN encourages latent dimensions to be mutually orthogonal and ordered by explained variance, mirroring the interpretable decomposition of PCA while retaining the expressive power of deep networks. We provide theoretical grounding for these constraints and demonstrate their compatibility with standard encoder-decoder frameworks. We also establish empirical results for both synthetic and real world datasets, establishing a principled path toward interpretable, structured feature learning and dimensionality reduction.

Time-domain surveys generate many transient candidates, making Real-Bogus classification a critical step in automated discovery pipelines. Reliable labels are costly, while community labels can be noisy and survey-dependent. We aim to develop a Real-Bogus classification framework that can be trained without human-labeled data using injected transients and bogus-dominated survey data, remains robust under strong class contamination, and provides calibrated uncertainty quantification. We combine simulated transient injections with a contaminated survey class and train a dual-network model using asymmetric co-teaching for classes with different label-noise levels. We evaluate performance on a benchmark subset and analyze the learned representation with latent-space visualization tools. For uncertainty quantification (UQ), we compare MC dropout and deep ensembles and propose a low-cost hybrid strategy that exploits the dual-network setting to improve calibration. We extend the evaluation to the light-curve domain to assess recovery of light-curve classes. The method achieves strong Real-Bogus performance on the labeled subset and remains stable under severe class contamination. It recovers transient light-curve classes with high fidelity, while single-source identification is limited by ambiguity in light-curve-derived labels. Our hybrid UQ approach achieves competitive calibration relative to more expensive ensemble baselines. Latent-space analyses indicate that uncertainty aligns with the decision boundary and reveal subclasses within the bogus population. Our results show that injection-driven, weakly supervised training can enable scalable and consistent Real-Bogus classification without human-labeled training data while providing calibrated uncertainties. The method is suited for transfer to forthcoming surveys by re-running the injection-based training pipeline.

3D reconstruction and generation are commonly tackled by separate paradigms: pixel-based regression for reconstruction, and latent diffusion for generation. Recent works attempt to unify them in latent space, but with notable drawbacks: the diffusion objective is defined on latent features rather than the underlying 3D representation, and both branches suffer from information loss introduced by latent encoding, while requiring a pretrained Variational Autoencoder (VAE) or Representation Autoencoder (RAE). In this paper, we reformulate these two tasks under a unified pixel-space diffusion paradigm and introduce PixWorld, a single model that jointly addresses 3D reconstruction and generation. By supervising diffusion directly on rendered images, PixWorld removes the above limitations and aligns optimization with 3D scene fidelity. Beyond photometric and perceptual supervision that operates at the 2D image level and lacks 3D geometric awareness, we further introduce a geometry perception loss that aligns rendered views with their ground truth in the geometry-aware feature space of a pretrained 3D foundation model, providing 3D structural supervision. PixWorld consistently outperforms prior latent-space generation methods and matches state-of-the-art reconstruction methods, demonstrating the superiority of a unified pixel-space approach.

We introduce the first multiplayer world model for highly dynamic environments governed by complex physical interactions. Whereas single-player world models treat the other agents as part of the environment, ours conditions on the action streams of multiple agents, learning to attribute changes in the scene to the correct player and to stay coherent under arbitrary combinations of their actions. We study this problem in the game of Rocket League, where players compete and cooperate under fast, tightly coupled dynamics. Trained on 10,000 hours of gameplay collected with publicly available bots, our 5-billion-parameter latent diffusion model generates four-player matches in real time, producing 20 frames per second on a single Nvidia B200 GPU. Although trained only on short clips, its rollouts stay stable far beyond the training horizon: distributional quality holds steady out to five minutes, the longest horizon we measure, and in practice we observe rollouts continuing for hours with no sign of collapse. We systematically investigate the central design choices: the video codec, the generative objective, and the multiplayer conditioning scheme. In addition, we characterize how behavior changes with model and data scale, including the capabilities that emerge and the failure modes that persist. We further develop targeted evaluations that probe the model's physical understanding rather than visual appearance alone. To support continued research on multiplayer world models, we release our dataset, our full training and inference codebase, and a live demo.

We study why diffusion autoencoders can achieve similar image quality while learning substantially different latent structures. We trace this behaviour to optimisation dynamics; we analyse curves of image reconstruction against latent representation quality, revealing trajectories that organise around two distinct regimes early in training. Models in the reconstruction regime prioritise image fidelity early, whereas those in the disentanglement regime improve reconstruction and disentanglement more gradually. We hypothesise that this behaviour can be influenced by targeting shortcut pathways in the diffusion U-Net and controlling early noise-level exposure, thereby shaping the reconstruction-disentanglement trade-off during training. To steer optimisation toward stronger representations, we introduce SteeringDRL, combining gated residual U-Nets with a simple noise-level exposure curriculum for training. Across disentanglement benchmarks, SteeringDRL improves representation quality and reduces seed sensitivity. Our method further extends to spatial disentanglement in object-centric learning, improving segmentation quality on synthetic and real-world datasets.

Learning motion latents for robotic manipulation heavily relies on extracting motion patterns from visual sequences, yet effective action abstractions require understanding three-dimensional geometric transformations. Here, we introduce GeoMoLa (Geometry-Aware Motion Latents), which learns discrete motion latent codes by predicting how point clouds evolve during manipulation rather than reconstructing visual observations. This four-dimensional objective -- spatial geometry changing through time -- forces latent representations to encode actual physical motion rather than appearance patterns. GeoMoLa achieves state-of-the-art performance using only single-view RGB-D input, while existing methods require multi-view reconstruction, succeeding across diverse manipulation benchmarks. Our ablations reveal that geometric prediction is the key to driving performance, quantitatively validating that manipulation depends on spatial understanding. Furthermore, the learned codes exhibit effective motion abstraction: applying them to novel scenes produces physically consistent transformations regardless of visual context. Our real-world experiments also confirm this robustness capability, achieving robust manipulation with minimal demonstrations in cluttered environments where geometric reasoning determines success. Thus, we demonstrate that effective motion latents for robot control can better emerge from understanding motion through its three-dimensional effects rather than pixel-level patterns.

We present Wan-Streamer v0.2, a latency-preserving upgrade of the native-streaming, end-to-end audio-visual interaction model. v0.2 keeps the v0.1 modeling formulation, but raises the interactive output stream from 192x336 to 640x368 while preserving approximately 200 ms model-side signal-to-signal latency at 25 FPS. The higher-resolution stream supports scene-grounded mid-shot agents whose posture, gaze, hands, nearby objects, and local scene layout remain legible during real-time conversation. To support the larger visual stream without adding user-visible delay, v0.2 keeps the thinker as a single-GPU low-latency path for streaming perception, the short language/state Transformer pass that builds the generation cache, and final decoding. The performer becomes a multi-GPU Ulysses-style context-parallel group for the expensive next-unit latent generation. Each performer rank writes incoming K/V into a pre-sharded local cache. The long high-resolution latent video sequence is split across ranks for denoising and gathered through Ulysses communication, while the much shorter audio latent sequence is generated without sequence sharding. In this split, the thinker's language/state computation reaches the performer only as K/V conditioning, so no separate language sequence has to be communicated inside the performer group. This concentrates additional hardware on visual generation while preserving the compact thinker-performer boundary, keeping total remote interaction latency at approximately 550 ms when a 350 ms bidirectional network budget is included.

Hierarchical structure is common in image data, where fine-grained clusters often merge into larger, coarser semantic groups. In biological cell images, current self-supervised learning models often suppress this hierarchy, as coarse factors such as imaging modality can obscure finer morphological attributes in the latent space. We propose a hierarchy-aware self-supervised training framework to address this problem. Our method combines two components: a distillation framework with a segmentation teacher to improve morphological awareness in the latent space, and a hierarchy-aware contrastive loss based on HDBSCAN to improve decision boundaries between closely related subtypes at different hierarchical levels. Together, these components reduce the tendency of self-supervised learning to overemphasize coarse factors and instead align embeddings with semantic and morphological cues. This yields biologically meaningful sub-clusters driven by fine morphological detail. We train and evaluate our method on a curated corpus of 2.3 million single cells aggregated from 20 microscopy datasets, both labeled and unlabeled, covering 208 cell classes. Our method improves over baseline and counterpart methods, increasing average top-K accuracy by 2.8%, top-9 retrieval on the dataset with the deepest hierarchy by 6.3%, and downstream F1-score for biologically relevant drug classification from perturbed cell morphology by 7.8%.

Precise control over complex dynamics remains challenging for modern video generative models, as text prompts alone often cannot specify physically plausible, fine-grained motion and interactions. We introduce $\textit{proxy-conditioned video generation}$, where a coarse proxy video from physics-based simulation or real-world recording serves as a dynamics carrier to control foreground object motion. Given a proxy video and a text prompt, the goal is to synthesize a new video that preserves the proxy dynamics while generating novel content and plausible interactions aligned with the prompt. Since paired proxy-target videos are difficult to obtain, we propose $\textbf{ProxyUp}$, a training-free framework built on pretrained video generative models. ProxyUp first inverts the proxy video into an intermediate latent representation and applies $\textbf{region-wise latent noising}$, preserving motion-critical proxy latents while injecting noise into regions intended for text-driven regeneration. To mitigate the distribution mismatch and weak foreground-background coupling introduced by this heuristic latent composition, we further propose $\textbf{Stochastic Flow Relaxation (SFR)}$, which progressively relaxes the composed latent toward the model's learned distribution before ODE sampling. Experiments on both simulation and real-world proxies show that ProxyUp outperforms strong video editing and motion transfer baselines in dynamic fidelity and text alignment.

The manifold hypothesis posits that high-dimensional data are concentrated near a low-dimensional embedded manifold. Recent advances in mixture variational autoencoders (VAEs) provide a powerful tool for extracting such underlying structure in a faithful manner. The resulting geometric structure naturally introduces local and global relationships among variables, thereby providing a systematic way of imputing missing data. We propose a model-based imputation method that enables sampling from $p(\bm{x}_{\mathrm{mis}} \mid \bm{x}_{\mathrm{obs}})$ via a sampling-importance-resampling (SIR) procedure, which can be further augmented with a joint diffusion model in the latent space. Our method imputes missing data while respecting the underlying geometry, achieves competitive performance compared to state-of-the-art procedures, quantifies uncertainty in the imputations, and is model-based, thereby enabling on-the-fly imputation without rerunning the entire procedure.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress but still struggle with complex visual reasoning tasks requiring multi-step perception and logical deduction. While explicit visual generation incurs prohibitive computational costs, existing latent approaches often rely on external experts or lack rigorous cognitive logic. In this paper, we introduce ProLaViT (Progressive Latent Visual Thought), a framework empowering MLLMs to perform structured visual derivation in the continuous latent space. Unlike works dependent on heterogeneous external models, ProLaViT leverages an endogenous self-distillation mechanism, utilizing the model's own visual encoder to supervise latent thoughts. To facilitate this, we construct a scalable programmatic synthesis pipeline enabling the model to internalize algorithmic precision without inference time tools. We design two reasoning paradigms: (1) Coarse-to-Fine Causal Chain for spatial tasks, guiding attention from global context to local targets. (2) Dialectical Reasoning Chain for logical tasks, incorporating counter-factual thinking for verification. Furthermore, we propose a Distance-Weighted Diversity Loss to impose topology-aware constraints, preventing feature degeneration by enforcing semantic distinctiveness. Extensive experiments demonstrate that ProLaViT outperforms baselines on vision-centric benchmarks, achieving superior accuracy and interpretability with high efficiency.

Realistic traffic simulation requires agents that imitate logged behavior and can also be steered along interpretable axes. Such controllability enables engineers to isolate variables, reproduce specific edge cases, and test autonomous systems without real-world risk. We introduce Controllable Neural Variational Agents (CNeVA), a controllable simulated-agent framework that learns to infer a per-agent Gaussian behavior latent from per-channel discounted returns via a closed-form conjugate variational update, conditioning a rectified-flow trajectory generator trained on a mixed channel-mask curriculum for classifier-free guidance. To tackle scarcity in reward signals, we propose soft eligibility gates that replace hard binary thresholds with smooth exponential decay, preserving the gradient signal for near-threshold agents. On the Waymo Open Motion Dataset, CNeVA attains competitive realism on the benchmark while exposing per-channel controllability that the higher-ranked imitation models lack. Speed- and acceleration-based steering produces monotone responses without stall-induced reward hacking. Safety controllability is monotone and substantial with the introduction of soft eligibility. We manage to achieve steerable map compliance under a context-residual return measure. Furthermore, our experiment demonstrates that steering metrics must be read alongside physical-plausibility guardrails to avoid reward-hacking confounds.