Video Latents

Visual Autoregressive (VAR) models deliver high-quality image generation but suffer from significant inference latency at high resolutions. Recent acceleration approaches most rely on heuristic measures with layer features to prune tokens. Such heuristics are sensitive to complex contextual semantics, leading to inaccurate identification of redundant computation and poor adaptability across prompts. We rethink redundancy in VAR from the perspective of its impact on pixel-space generation and introduce Latent Discrepancy. This unified metric quantifies a token's contribution by measuring the change in model states during generation. Our analysis shows that redundancy is more accurately identified when guided by image latent or pixel-space signals. We further observed that in classifier-free guidance (CFG), the convergence trend of the discrepancy between conditional and unconditional branches exhibits high dynamics with different prompts. Based on these findings, we propose LD-Pruning (Latent Discrepancy Pruning), a training-free framework that removes redundancy via latent discrepancy by integrating decoding-free region selection and adaptive unconditional-branch skipping. Extensive experiments show that LD-Pruning substantially reduces inference latency while maintaining high generation quality, achieving up to 2.35x speedup on Infinity-8B.

Frame-wise action-controlled image-to-video generation is a promising paradigm for interactive world simulation, where each control signal should elicit an immediate visual response. However, maintaining visual fidelity and 3D consistency over long autoregressive rollouts remains challenging. Existing 3D-aware methods often suffer from catastrophic drift due to two impediments: information loss from \textit{Latent--RGB Cycling}, where generated latents are repeatedly decoded to RGB and re-encoded for future conditioning, and the training--inference gap induced by the \textit{error-free hypothesis}, where clean training memory fails to match prediction-corrupted inference memory. To address these challenges, we present \textbf{Robust Dreamer}, a memory-augmented framework built around how to design 3D memory and how to use it robustly. First, we introduce \textbf{Latent Gaussian Memory}, which anchors diffusion latents inherited from the generation process to Gaussian primitives and recalls them via latent-space Gaussian splatting. This provides dense, geometry-aware, view-aligned conditioning while avoiding accumulated degradation from repeated VAE conversion. Second, we propose \textbf{Deviation Learning with Dynamic Deviation Archive}, which synthesizes rollout-induced latent deviations through a one-step approximation, stores them by autoregressive stage and denoising timestamp, and injects them into historical memory during training. This exposes the generator to realistic corrupted memory states and teaches internal correction before inference. Experiments on ScanNet, DL3DV, and OmniWorldGame demonstrate state-of-the-art long-horizon performance.

Critical networking workflows require high-fidelity packet captures (PCAPs) for testing, security analysis, and protocol validation, not just statistical flow-level summaries. Recent packet generators have demonstrated protocol-constrained PCAP synthesis, but they universally decode directly to raw packet fields. That interface entangles learned behavioral choices with deterministic protocol consequences, which forces packet realization to depend on post-hoc heuristic repair. We identify this decode interface as the fundamental bottleneck and present TraceCodec, a state-aware neural codec for stateful multi-flow traces. TraceCodec lifts each packet into a timed packet action with explicit flow slots and transport cues, then learns a continuous per-packet latent. A deterministic compiler lowers decoded actions back to PCAPs, owning endpoint assignment, TCP state, legality constraints, and packet rendering. The latent layer exposes a generator-facing sequence space, so downstream traffic models can operate on packet-action latents rather than raw header fields. On CICIDS2017 Monday, TraceCodec matches packet count, protocol composition, and flow population to within 0.03%. Raw-field baselines under the same non-repair policy distort flow counts and TCP state by orders of magnitude. Structural diagnostics show that TraceCodec preserves TCP state transitions and multi-flow interleaving that raw-field decoders fragment. This work establishes a new foundation for high-fidelity packet-trace generation.

Autoregressive multimodal large language models (MLLMs) enable 3D generation but struggle to scale to high-resolution shapes due to inadequate 3D tokenizations. Compact set-based representations discard deterministic spatial ordering, leading to ambiguous sequence prediction, while uniform or octree-based voxel grids preserve ordering at the cost of severe redundancy and excessively long sequences. This structural trade-off limits stable and efficient autoregressive 3D generation. We present SuperVoxelGPT, a representation-first framework that resolves this tension through adaptive and deterministically ordered supervoxel tokenization. Given a prompt, we first predict a coarse geometric saliency distribution and construct a shape-adaptive supervoxel partition using saliency-guided centroidal Voronoi tessellation, allocating fine-grained cells to complex regions and larger cells to smooth regions. Conditioned on this prompt and ordered supervoxel layout, we introduce a SuperVoxelVAE and fine-tune a pretrained MLLM to autoregressively generate supervoxel tokens. Experiments using Trellis-500K data show that SuperVoxelGPT reduces token sequence length to 12.8% of uniform voxel tokenization while achieving state-of-the-art generation quality and an average 10x speedup over prior methods.

Recent approaches integrating vision-language models (VLMs) as prompt encoders for generative model conditioning typically rely on expensive end-to-end training or map features to compressed representations, discarding the dense spatial structure required for geometry-aware tasks like 3D asset generation. To address this, we propose GAP3D, a modular, diffusion-based approach that aligns VLM-generated latents directly to the complete, patch-level feature space of a pre-trained image encoder, enabling a frozen downstream generative model to utilize a VLM as prompt encoder while maintaining a spatially structured conditioning signal. Evaluated on 3D asset generation, our method bypasses the need for large-scale 3D data by training mainly on general-domain image-text pairs. It also exhibits emergent zero-shot capabilities for multimodal prompts, despite being trained exclusively on text input. Finally, while currently prioritizing high-level semantics over fine-grained detail, GAP3D demonstrates that the representation gap between VLM and image-encoder feature spaces can be partially bridged through diffusion-based alignment, taking the first steps towards a modular integration of foundation models through generative alignment to dense embedding spaces.

Sparse Autoencoders (SAEs) extract interpretable features from Large Language Models, but standard variants enforce non-negativity, forcing separate latents for diametrically opposed concepts (e.g., "pressure too high" vs. "pressure too low") and wasting dictionary capacity when features are anticorrelated. We propose the Sign-Aware Gated SAE (SA-GSAE): two-sided gated sparsity with signed magnitude and auxiliary supervision. A polarity-sensitive gate selects support on either sign, a signed-magnitude path avoids L1 shrinkage, and an auxiliary reconstruction prevents gate collapse. Bipolar sharing - one latent encoding both signs along a shared direction - is realised via a new Bi-Jump-ReLU activation; parameter accounting shows sign-awareness stays parameter-efficient even when anticorrelated pairs are rare. On real LLM activations across three mid-depth hookpoints on Pythia-1B and SmolLM3-3B (6 cells, 3 seeds), a half-width SA-GSAE at width H strictly Pareto-dominates a full-width Gated SAE at 2H over the entire swept L0 overlap on 3 of 6 cells (both MLP-output hookpoints and resid-mid/Pythia-1B); on the remaining 3 it matches R^2 within 0.025 (max gap -0.008) while cutting dead fraction by 0.35-0.62 absolute. Sweep-geomean dead-fraction reductions are ~100x-500x on MLP-output cells and Pythia-1B resid, ~2x-4x on attention cells and SmolLM3-3B resid. Ablations show the two-sided gate and auxiliary loss are load-bearing (no auxiliary collapses LR to 0.27, 98% dead); tying r_i^+ = r_i^- is indistinguishable (|Delta R^2| = 0.0015), and we recommend this symmetric variant as default. MLP-output gains come from most latents carrying both polarities; on attention, bipolar structure concentrates in a small set of top latents. Full-width SA-GSAE exhibits a reproducible reconstruction collapse at SmolLM3-3B resid that the half-width entirely avoids.

Vision-Language-Action (VLA) models commonly adapt pretrained Vision-Language Models (VLMs) to robot control by mapping visual observations and language instructions to continuous actions. Existing approaches typically take an action-insufficiency view, assuming that pretrained VLM latents either lack directly usable action information or should be shielded from action-learning signals. Against this view, our \textit{Quotient Theory for VLA} shows that pretrained VLM latents are not action-insufficient but action-sufficient: they already contain the information needed for control, yet remain overcomplete by distinguishing prompt-level variations that induce the same optimal action behavior. To operationalize this theory, we propose QuoVLA, a quotient-space framework for VLA that compresses pretrained VLM latents into action-sufficient representations. Specifically, QuoVLA instantiates this principle with a quantization module and a dual-branch design with relative temporal-complexity regularization, preserving action-relevant information while removing prompt-level redundancy. Extensive experiments across multiple benchmarks demonstrate that QuoVLA achieves strong performance, with particularly notable improvements in generalization under visual, linguistic, and environmental distribution shifts. Our code will be made publicly available.

Pose-driven full-body avatars built on neural rendering produce high-quality novel views of a captured subject. Yet loose clothing and other dynamic elements deform in ways pose alone cannot explain: the same pose can correspond to many different states, because their motion depends on history, inertia, and contact. Explicit simulation and layered-garment methods can model such dynamics, but they require either a dedicated garment template, which raw multi-view capture does not naturally provide, or a test-time physics simulator with non-trivial runtime cost. A parallel line of work learns data-driven clothing avatars that avoid explicit garment layers. These methods add an auxiliary latent for variation beyond pose; at inference, they fix it, regress it from pose, or retrieve it from training data, without explicitly modeling how the latent evolves with its own dynamics. Additionally, even in everyday motion with loose clothing, existing architectures often struggle to capture fine-grained detail, producing blurry renderings and temporal artifacts. We augment a pose-conditioned 3D Gaussian avatar with a transformer-based decoder and a dynamics residual latent that captures temporal appearance and geometry variation beyond the driving signals. At inference, a learned latent dynamics model evolves the residual latent from a short pose history and the previous latent state. The model decomposes each update into driving, restoring, and dissipative forces, producing temporally coherent, history-dependent rollouts with negligible added cost. Different initial conditions yield diverse yet plausible motion trajectories, and the force decomposition exposes controls such as stiffness. Across nine captured sequences of everyday motion with diverse loose garments, quantitative metrics and a perceptual user study show improved animation quality over recent data-driven baselines.

Single-image-to-3D generative models can now produce high-quality geometry, yet conditioning on a single view inevitably introduces ambiguity about unseen regions. Multi-view conditioning can reduce this ambiguity, but existing methods either require fixed canonical viewpoints or rely on external reconstruction modules that impose heavy training costs and limit generation quality. We observe that pretrained single-view models already possess strong 2D-to-3D grounding that can be reused for multi-view conditioning. However, a closer analysis reveals that their conditioning mechanism entangles orientation control with geometry transfer, two functions that conflict when images from different viewpoints are naively combined. Based on this analysis, we propose ROAR-3D, a lightweight method that upgrades a pretrained single-view model to accept an arbitrary number of unposed images. A token-wise view router assigns each 3D latent token to its most relevant view, implicitly establishing 2D-to-3D correspondences without explicit pose input. A dual-stream attention design preserves the pretrained primary-view behavior while routing auxiliary views through a separate path dedicated to geometric enrichment. An orientation perturbation strategy ensures the auxiliary path learns orientation-independent geometry transfer. These components introduce minimal trainable parameters and add negligible inference overhead relative to the single-view baseline. ROAR-3D achieves state-of-the-art multi-view 3D generation quality and supports test-time view scaling from 1 to 12+ views with consistent improvements.

Multimodal latent-space reasoning aims to replace explicit thinking with images by performing visual reasoning directly in a compact latent space. However, existing approaches largely rely on visual supervision and produce latent representations that lack sufficient semantic richness, limiting their ability to support diverse region-level reasoning tasks. In this work, we introduce Semantic-Enriched Latent Visual Reasoning (SLVR), a two-stage learning framework that enriches latent representations with attribute-level visual semantics and aligns them with diverse reasoning objectives. In the first stage, SLVR learns semantically enriched region-centric latents under fine-grained attribute supervision. In the second stage, we design Multi-query Group Relative Policy Optimization (M-GRPO) to align latent representations across multiple queries grounded in the same region. To support this framework, we construct SLV-Set, comprising approximately 400K region-level attribute annotations and 800K multi-query question answering samples, and introduce SV-QA, a benchmark that evaluates latent reasoning under semantic variation. Experiments demonstrate that SLVR improves the robustness and semantic consistency of latent visual reasoning compared to existing baselines.

Vision-Language-Action (VLA) models leverage powerful perceptual priors from web-scale Vision-Language Model (VLM) pre-training, yet they remain surprisingly brittle in practice, frequently failing at simple robotic tasks. To mitigate this, we propose Contrastive Conceptor Activation Steering (COAST). COAST builds on the notion of a "conceptor", a linear operator that soft-projects data into the principal components of a target distribution. COAST uses conceptors to identify success-critical subspaces for a target robotic task from a few examples of success and failure rollouts. At inference time, it steers VLA latents into these identified success subspaces to improve task outcomes. Across three architecturally distinct neural policies (flow-matching VLA, autoregressive VLA, and Diffusion Policy), COAST improves absolute mean simulation and real-robot task success rate by over 20 and 40% respectively. The activation subspace geometry reveals that failure modes share substantial structure across tasks while success representations remain largely task-specific. When tasks share similar failure modes, this structure enables previously fitted conceptors to improve performance on new tasks without refitting. Ultimately, our results suggest that current VLAs retain substantial task-relevant knowledge in their latent representations, and that the action expert's decoding bottleneck could be mitigated by steering its residual stream toward task-relevant subspaces. COAST provides a lightweight, training-free path to unlocking these latent capabilities by steering the model towards its own "success" distributions.

Large language models (LLMs) have recently been used as structured decoders for indoor understanding from 3D point-token inputs. However, point cloud encoders often under-represent thin structural elements such as doors and windows after voxelization and sparse pooling, and may miss individual furniture instances in cluttered scenes. We propose an interface-preserving semantic augmentation for LLM-conditioned structured decoding. The key idea is to associate semantic evidence with the point-cloud representation, reduce it to a coarse four-group code (furniture, walls, openings, and others), and encode it as an RGBB point interface: red for furniture, green for walls, blue for openings, and black for others, where RGBB denotes four semantic color states represented in three RGB channels rather than an additional fourth channel. This semantic color code is appended to the original raw point attributes before tokenization, so geometry and semantics share the same sparse tokenization path while the downstream language model decoder and output serialization remain unchanged. We further introduce a lightweight routed semantic shift module, with an auxiliary head used only for training-time ratio/budget regularization and analysis, to strengthen semantic cues after sparse pooling. The overall pipeline can use RGB-derived semantic evidence. Under these controlled semantic-source settings, the reported metrics improve across Structured3D, the SpatialLM dataset, and ARKitScenes, especially for opening localization and per-instance furniture detection in cluttered scenes. Ablations clarify the roles of semantic source, color coding, token fusion, and shift injection, while also showing that color/entropy effects remain nontrivial.

Natural images are continuous, yet most generative models synthesize them on discrete grids, limiting resolution-flexible generation. Continuous neural fields enable resolution-free rendering, but prior methods introduce continuity only at the decoding stage as an interpolation module, leaving the generative latent space discretized and reconstruction-oriented. We propose RaPD (Resolution-agnostic Pixel Diffusion), which performs diffusion in a continuous Neural Image Field (NIF) latent space. RaPD bridges this reconstruction-generation gap with Semantic Representation Guidance for generation-aware latent learning and a Coordinate-Queried Attention Renderer for coordinate-conditioned, scale-aware rendering. A single denoised latent can be rendered at arbitrary resolutions by changing only the query coordinates, keeping diffusion cost fixed. Experiments demonstrate superior generation quality and resolution scalability.

Continuous diffusion and flow models are attractive for non-autoregressive text generation because they can update all positions in parallel. A major difficulty is the interface between continuous latent states and discrete tokens. This report studies a draft-conditioned latent refinement model built from a frozen BERT encoder, a parallel decoder, a denoising DraftPrior, a local FlowNet, and a learned diagonal MetricNet. Early Gaussian-start experiments showed that good latent-space metrics, such as scale matching or cosine similarity, do not guarantee good decoding. Generated latents can be close to real encoder latents but still produce high-entropy, biased, or repetitive token distributions. We therefore frame the task as controlled local refinement rather than full generation from noise. On ROCStories, using the first two sentences as prompt and the last three as target, full 768-dimensional BERT latents recover tokens much better than compressed 256-dimensional latents. With 768-dimensional latents, DraftPrior target-token probability is 0.938 for clean drafts, 0.613 for 3% token dropout, 0.483 for 5% dropout, and 0.272 for 10% dropout. Local flow refinement and fused decoder-aware readout give modest additional gains, while metric learning and OT-style alignment improve geometry but do not close the decoder gap. The main result is a diagnostic one: latent geometry alone is not enough. Continuous latent text generation should be evaluated by decoder recoverability, the quality of the start distribution, and whether refinement preserves decoder-readable structure.

We propose EverAnimate, an efficient post-training method for long-horizon animated video generation that preserves visual quality and character identity. Long-form animation remains challenging because highly dynamic human motion must be synthesized against relatively static environments, making chunk-based generation prone to accumulated drift: (i) low-level quality drift, such as progressive degradation of static backgrounds, and (ii) high-level semantic drift, such as inconsistent character identity and view-dependent attributes. To address this issue, EverAnimate restores drifted flow trajectories by anchoring generation to a persistent latent context memory, consisting of two complementary mechanisms. (i) Persistent Latent Propagation maintains a context memory across chunks to propagate identity and motion in latent space while mitigating temporal forgetting. (ii) Restorative Flow Matching introduces an implicit restoration objective during sampling through velocity adjustment, improving within-chunk fidelity. With only lightweight LoRA tuning, EverAnimate outperforms state-of-the-art long-animation methods in both short- and long-horizon settings: at 10 seconds, it improves PSNR/SSIM by 8%/7% and reduces LPIPS/FID by 22%/11%; at 90 seconds, the gains increase to 15%/15% and 32%/27%, respectively.

Text-driven motion diffusion models are capable of generating realistic human motions, but text alone often struggles to express fine-level nuances of motion, commonly referred to as style. Recent approaches have tackled this challenge by attaching a style injection mechanism to a pretrained text-driven diffusion model. Existing stylization methods, however, either require style-specific fine-tuning of existing models or rely on heavy ControlNet-based architectures, limiting efficiency and generalization to unseen styles. We propose a lightweight style conditioning framework that dynamically modulates a pretrained diffusion model through hypernetwork-generated LoRA parameters. A style reference motion is encoded into a global style embedding, which is mapped by a hypernetwork to low-rank updates applied at each denoising step of the diffusion model. By structuring the style latent space with a supervised contrastive loss, our framework reliably captures diverse stylistic attributes, improves generalization to unseen styles, and supports optimization-based guidance without requiring predefined style categories. Experiments on the HumanML3D and 100STYLE datasets show state-of-the-art stylization results, while achieving improved stylization for unseen styles.

Recent 3D generative models can synthesize high-quality assets, but their outputs are typically static: they lack the skeletal rigs, joint hierarchies, and skinning weights required for animation. This limits their use in games, film, simulation, virtual agents, and embodied AI, where assets must not only look plausible but also move plausibly. We introduce Rigel3D, a generative method for animation-ready 3D assets represented as rigged meshes. Unlike post-hoc auto-rigging methods that attach rigs to completed shapes, our method jointly models geometry and rig structure through coupled surface and skeleton structured latent representations. A rig-aware autoencoder decodes these representations into mesh geometry, skeleton topology, joint coordinates, and skinning weights, while a two-stage latent generative model synthesizes both surface and skeleton representations for image-conditioned generation. To support downstream animation workflows, we further introduce an open-vocabulary joint labeling module that embeds generated joints into a shared vision-language space, enabling correspondence to arbitrary retargeting templates. Experiments on large-scale rigged asset datasets demonstrate that our method generates diverse, high-quality animation-ready assets and outperforms existing rigging baselines across multiple metrics.

Dense 3D tracking from monocular video is fundamental to dynamic scene understanding. While recent 3D foundation models provide reliable per-frame geometry, recovering object motion in this geometry remains challenging and benefits from strong motion priors learned from real-world videos. Existing 3D trackers either follow iterative paradigms trained from scratch on synthetic data or fine-tune 3D reconstruction models learned from static multi-view images, both lacking real-world motion priors. Pre-trained video diffusion transformers (video DiTs) offer rich spatio-temporal priors from internet-scale videos, making them a promising foundation for 3D tracking. However, their frame-anchored formulation, which generates each frame's content, is fundamentally mismatched with reference-anchored dense 3D tracking, which must follow the same physical points from a reference frame across time. We present TrackCraft3R, the first method to repurpose a video DiT as a feed-forward dense 3D tracker. Given a monocular video and its frame-anchored reconstruction pointmap, TrackCraft3R predicts a reference-anchored tracking pointmap that follows every pixel of the first frame across time in a single forward pass, along with its visibility. We achieve this through two designs: (i) a dual-latent representation that uses per-frame geometry latents and reference-anchored track latents as dense queries, and (ii) temporal RoPE alignment, which specifies the target timestamp of each track latent. Together, these designs convert the per-frame generative paradigm of video DiTs into a reference-anchored tracking formulation with LoRA fine-tuning. TrackCraft3R achieves state-of-the-art performance on standard sparse and dense 3D tracking benchmarks, while running 1.3x faster and using 4.6x less peak memory than the strongest prior method. We further demonstrate robustness to large motions and long videos.

The field of image-to-video generation has made remarkable progress. However, challenges such as human limb twisting and facial distortion persist, especially when generating long videos or modeling intensive motions. Existing human image animation works address these issues by incorporating human-specific semantic representations, e.g., dense poses or ID embeddings, as additional conditions. However, conditioning on these representations could decrease the generation flexibility. Moreover, their reliance on RGB pixel supervision also lacks emphasis on learning necessary 3D geometric relationships and temporal coherence. In contrast, we introduce a novel approach named SemanticREPA that leverages these semantic representations as supervision signals through representation alignment. Specifically, we begin by training a structure alignment module that aligns the structure representations obtained from video latents with video depth estimation features. We then fix the pretrained module, and utilize it to provide additional supervision on the structure representations of the diffusion models, achieving structure rectification to generate coherent and stable human structures. Simultaneously, we develop an ID alignment module to align the ID representations of the generated videos to face recognition features. We further propose to use the predicted structure representations to refine identity restoration in relevant regions. With structure and ID alignment, our method demonstrates superior quality on extended character motions and enhanced character consistency.

Many features in pretrained Transformers span multiple layers: they emerge through stages of inference, persist in the residual stream, or are built jointly by parallel MLPs. Crosscoders (namely, sparse dictionaries trained jointly across layers) aim to recover these cross-layer features in a single shared latent space. We show that standard crosscoders largely fail at this purpose. Although their decoder weight norms spread evenly across layers, a functional coherence metric we introduce reveals that each latent's activation is effectively driven by only one or two layers on average. While functionally coherent latents act as human-interpretable concept detectors (e.g., US states and cities), the layer-localized latents that crosscoders predominantly learn collapse onto surface-level patterns such as digit detectors. We trace this failure to two structural limitations: unconstrained cross-layer parameterization and unregularized cross-layer dependence. We address both by introducing fmxcoders, which (i) replace the encoder and decoder with low-rank tensor factorizations that draw every latent's per-layer weights from a shared cross-layer basis, and (ii) apply stochastic layer masking, a denoising regularizer along the layer axis that penalizes latents whose contribution collapses when a single layer is masked. Across GPT2-Small, Pythia-410M, Pythia-1.4B, and Gemma2-2B, fmxcoders lift mean probing F1 by 10-30 points, surpassing per-layer SAE baselines that standard crosscoders fail to reach, reduce reconstruction MSE by 25-50%, and roughly double mean functional coherence. An LLM-as-a-judge evaluation further shows that fmxcoders recover 3-13$\times$ more semantically coherent latents than standard crosscoders across all four base LLMs.

This paper investigates robust representation learning in offline goal-conditioned reinforcement learning (GCRL). Particularly in sparse reward scenarios, learning representations that align state and goal latents is a challenge that frequently culminates in representation divergence where the encoder drifts toward a low-dimensional, goal-agnostic subspace that destabilizes policy learning. We address this issue by showing that an agent must acquire a fundamental understanding of its environment across multiple scales, from local physical dynamics to long-horizon goal-directed structure. Building on this insight, we propose Ms.PR, a framework that leverages multi-scale predictive supervision to enforce goal-directed alignment within the latent space. We demonstrate that Ms.PR leads to improved representation quality and strong performance on both vision and state-based tasks. Furthermore, we show that our approach is exceptionally resilient under realistic, challenging data regimes, maintaining state-of-the-art performance across a wide variety of tasks, trajectory stitching scenarios, and extreme noise conditions.

With the proliferation of AI-generated images, digital watermarking has become an essential safeguard for protecting intellectual property and mitigating malicious exploitation. Recent works on semantic watermarking have enabled efficient copyright protection for diffusion models. However, the dependence of semantic watermarking on diffusion inversion for watermark detection creates a critical vulnerability. Imprint removal and forgery attacks exploit this weakness to produce deceptive results. Our analysis reveals that these attacks succeed by displacing watermarked latents into the unwatermarked region, while guiding unwatermarked latents into the watermarked region. Based on that, we propose Progressive Guided Inversion and Denoising (PGID), the first plug-and-play, training-free noise extraction framework designed to defend against both attack strategies. PGID effectively defends by projecting perturbed latents back to the region where they originally belong. The projection is achieved by eliminating intermediate latent deflections and mitigating adversarial perturbations through progressive inversion-denoising cycles. Comprehensive evaluations across multiple schemes demonstrate that PGID successfully restores detection reliability by recovering removed watermarks and identifying forged instances.

Sparse autoencoders have become a standard tool for uncovering interpretable latent representations in neural networks. Yet salient concepts often span manifolds that current linear methods cannot capture without post hoc analysis. This paper uses quadratic latents to close this gap: we implement these with bilinear autoencoders, which decompose activations into low-rank quadratic forms, compose linearly in weight space, and admit input-independent geometric analysis. This qualitative difference in what concepts quadratic latents can detect challenges the standard linear representation hypothesis. Our experiments and visualisations show that multi-dimensional geometries are highly prevalent and that composite latents capture them well, systematically improving reconstruction error in language models. Furthermore, we show that autoencoders with varying geometric priors recover the same input subspace despite their dictionary entries being distinct. Practically, these models serve as an unsupervised tool for manifold discovery, which we demonstrate through an interactive online visualizer for Qwen 3.5. This is a step toward nonlinear but mathematically tractable latent representations whose composition is expressive and interpretable by design.

We present the Wristband Gaussian Loss, a deterministic batch loss for Gaussianizing point embeddings without sampling, KL terms, or iterative transport. Each $x \in \mathbb{R}^d$ is mapped to a direction $u = x/\|x\|$ and a CDF-transformed radius $t = F_{χ^2_d}(\|x\|^2)$ on the wristband $S^{d-1} \times [0,1]$. We prove (and machine-verify in Lean~4) that for $d \ge 2$ the pushforward wristband map equals $σ_{d-1} \otimes \mathrm{Unif}[0,1]$ iff the source is $\mathcal{N}(0, I_d)$, and that the Neumann-reflected wristband repulsion energy is uniquely minimized at the uniform target. We compute this reflected-kernel objective in two ways: a nearest three-image pairwise truncation at $O(N^2 d)$, and a spectral Neumann path joining angular and radial Mercer modes (spherical-harmonic and cosine) at $O(N d K)$, with empirically matched gradients. A 1D Wasserstein radial term and a moment penalty serve as finite-sample accelerators with the same optimum, and Monte-Carlo null calibration turns the components into a single standardized statistic. We evaluate direct point-cloud Gaussianization with a calibrated barycentric $W_2$ score: a deterministic Gaussian reference batch is built by recursive Hungarian averaging, with each method reported as a $z$-score against same-size Gaussian batches. On the axis-uniform X benchmark, Wristband is competitive in 2D and gives the best 10D score. On a harder radial--angular-copula impostor whose Gaussian radial and angular marginals are correct but dependent, Wristband gives the best 10D and 128D scores. Coupled with learnable-key Euclidean attention and exact invertible flows, the resulting Deterministic Gaussian Autoencoder delivers a Gaussian-latent interface for counterfactual sampling with independent factors and a context/residual construction for dependent factors.

Large language models have achieved remarkable success under the autoregressive paradigm, yet high-quality text generation need not be tied to a fixed left-to-right order. Existing alternatives still struggle to jointly achieve generation efficiency, scalable representation learning, and effective global semantic modeling. We propose Cola DLM, a hierarchical latent diffusion language model that frames text generation through hierarchical information decomposition. Cola DLM first learns a stable text-to-latent mapping with a Text VAE, then models a global semantic prior in continuous latent space with a block-causal DiT, and finally generates text through conditional decoding. From a unified Markov-path perspective, its diffusion process performs latent prior transport rather than token-level observation recovery, thereby separating global semantic organization from local textual realization. This design yields a more flexible non-autoregressive inductive bias, supports semantic compression and prior fitting in continuous space, and naturally extends to other continuous modalities. Through experiments spanning 4 research questions, 8 benchmarks, strictly matched ~2B-parameter autoregressive and LLaDA baselines, and scaling curves up to about 2000 EFLOPs, we identify an effective overall configuration of Cola DLM and verify its strong scaling behavior for text generation. Taken together, the results establish hierarchical continuous latent prior modeling as a principled alternative to strictly token-level language modeling, where generation quality and scaling behavior may better reflect model capability than likelihood, while also suggesting a concrete path toward unified modeling across discrete text and continuous modalities.

Training world models on vast quantities of unlabelled videos is a critical step toward fully autonomous intelligence. However, the prevailing paradigm of encoding raw pixels into opaque latent spaces and relying on heavy decoders for reconstruction leaves these models computationally expensive and uninterpretable. We address this problem by introducing NOVA, a world modelling framework that represents the system state as the weights and biases of an auxiliary coordinate-based implicit neural representation (INR). This structured representation is analytically rendered, which eliminates the decoder bottleneck while conferring compactness, portability, and zero-shot super-resolution. Furthermore, like most latent action models, NOVA can be distilled into a context-dependent video generator via an action-matching objective. Surprisingly, without resorting to auxiliary losses or adversarial objectives, NOVA can disentangle structural scene components such as background, foreground, and inter-frame motion, enabling users to edit either content or dynamics without compromising the other. We validate our framework on several challenging datasets, achieving strong controllable forecasting while operating on a single consumer GPU at $\sim$40M parameters. Ultimately, structured representations like INRs not only enhance our understanding of latent dynamics but also pave the way for immersive and customisable virtual experiences.

Dynamic spatial reasoning from monocular video is essential for bridging visual intelligence and the physical world, yet remains challenging for vision-language models (VLMs). Prior approaches either verbalize spatial-temporal reasoning entirely as text, which is inherently verbose and imprecise for complex dynamics, or rely on external geometric modules that increase inference complexity without fostering intrinsic model capability. In this paper, we present 4DThinker, the first framework that enables VLMs to "think with 4D" through dynamic latent mental imagery, i.e., internally simulating how scenes evolve within the continuous hidden space. Specifically, we first introduce a scalable, annotation-free data generation pipeline that synthesizes 4D reasoning data from raw videos. We then propose Dynamic-Imagery Fine-Tuning (DIFT), which jointly supervises textual tokens and 4D latents to ground the model in dynamic visual semantics. Building on this, 4D Reinforcement Learning (4DRL) further tackles complex reasoning tasks via outcome-based rewards, restricting policy gradients to text tokens to ensure stable optimization. Extensive experiments across multiple dynamic spatial reasoning benchmarks demonstrate that 4DThinker consistently outperforms strong baselines and offers a new perspective toward 4D reasoning in VLMs. Our code is available at https://github.com/zhangquanchen/4DThinker.

3D asset generation plays a pivotal role in fields such as gaming and virtual reality, enabling the rapid synthesis of high-fidelity 3D objects from a single or multiple images. Building on this capability, enabling style-controllable generation naturally emerges as an important and desirable direction. However, existing approaches typically rely on style images that lie within or are similar to the training distribution of 3D generation models. When presented with out-of-distribution (OOD) styles, their performance degrades significantly or even fails. To address this limitation, we introduce \textbf{DiLAST}: 2D Diffusion-based Latent Awakening for 3D Style Transfer. Specifically, we leverage a pretrained 2D diffusion model as a teacher to provide rich and generalizable style priors. By aligning rendered views with the target style under diffusion-based guidance, our method optimizes the structured 3D latent representations for stylization. We observe that this limitation stems not from insufficient model capacity, but from the underutilization of structured 3D latents, which are inherently expressive. Despite being trained on comparatively limited data, 3D generation models can leverage 2D diffusion guidance to steer denoising toward specific directions in latent space, thereby producing diverse, OOD styles. Extensive experiments across diverse data and multiple 3D generation backbones demonstrate the effectiveness and plug-and-play nature of our approach.

Continuous latent-space reasoning offers a compact alternative to textual chain-of-thought for multimodal models, enabling high-dimensional visual evidence to be integrated without explicit reasoning tokens. However, we identify a previously overlooked optimization pathology in existing latent visual reasoning methods: although visual latents become semantically enriched during training, their contribution to final answer prediction is systematically suppressed. Within the shared parameter space, the autoregressive objective favors shortcut reliance on direct visual input, driving latent tokens toward transition-like states rather than informative reasoning content. We term this phenomenon Silenced Visual Latents. To address it, we disentangle the two conflicting objectives by directly optimizing the latent reasoning at inference time, keeping backbone parameters frozen. In Stage I, visual latents are warmed up via query-guided contrastive latent--visual alignment, improving semantic quality while preventing latent collapse. In Stage II, the latent reasoning is further optimized via a confidence-progression reward, which incentivizes predicted token distributions along the latent span to become progressively more concentrated, routing predictions through the latent reasoning rather than bypassing it. Experiments across eight benchmarks and four model backbones show that inference-time latent optimization, without any parameter updates, effectively unleashes the suppressed reasoning capacity of visual latents.

We present a training-free approach for controllable 3D inpainting based on initial noise optimization. In the structured 3D latent diffusion framework, we observe that the underlying geometric structure is established during the early stages of the diffusion process and exhibits high sensitivity to the initial noise. Such characteristics compromise stability in tasks like inpainting and editing, where the model must ensure strict alignment with the existing context while synthesizing a new structure. In this paper, we introduce a strategy to optimize the initial noise within the structured 3D latent diffusion framework, ensuring high-fidelity 3D inpainting. Specifically, we update the initial noise by leveraging a backpropagation approximation grounded in the rectified flow model, with the spectral parameterization specially designed for robust and efficient structured 3D latent optimization. Experiments demonstrate consistent improvements in contextual consistency and prompt alignment over representative training-free inpainting baselines, establishing initial noise control as an independent dimension for 3D inpainting, orthogonal to conventional sampling trajectory manipulation.

Video generation, while capable of generating realistic videos, is computationally expensive and slow, prohibiting real-time applications. In this paper, we observe that video latents encoded via an autoencoder under the Latent Diffusion Model (LDM) framework contain redundancy along the temporal axis. Analogous to how traditional video compression algorithms avoid transmitting redundant frame data, we propose the Latent Inter-frame Pruning framework to prune (skip the re-computation of) duplicated latent patches, thereby reducing computational burden and increasing throughput. However, direct pruning results in visual artifacts due to the discrepancy between full-sequence training and pruned inference. To resolve these artifacts, we propose an Attention Recovery mechanism to bridge the train-inference gap. With our proposed method, we increase video editing throughput by 1.44$\times$, achieving 12.44 FPS on an NVIDIA RTX 6000 while maintaining video quality. We hope our work inspires further research into integrating traditional video compression methods with modern video generation pipelines. This work is a preliminary work on Training-free Latent Inter-Frame Pruning with Attention Recovery.

Models from the AlphaFold (AF) family reliably predict one dominant conformation for most well-ordered proteins but struggle to capture biologically relevant alternate states. Several efforts have focused on eliciting greater conformational variability through ad hoc inference-time perturbations of AF models or their inputs. Despite their progress, these approaches remain inefficient and fail to consistently recover major conformational modes. Here, we investigate both the optimal location and manner-of-operation for perturbing latent representations in the AF3 architecture. We distill our findings in ConforNets: channel-wise affine transforms of the pre-Pairformer pair latents. Unlike previous methods, ConforNets globally modulate AF3 representations, making them reusable across proteins. On unsupervised generation of alternate states, ConforNets achieve state-of-the-art success rates on all existing multi-state benchmarks. On the novel supervised task of conformational transfer, ConforNets trained on one source protein can induce a conserved conformational change across a protein family. Collectively, these results introduce a mechanism for conformational control in AF3-based models.

Industrial CAD workflows require robust, generalizable 3D geometric representations supporting accuracy and explainability. We introduce Shape, a self-supervised foundation model converting surface meshes into dense per-token embeddings. Shape combines a structured 3D latent grid, a multi-scale geometry-aware tokenizer (MAGNO) with cross-attention, and a transformer processor using grouped-query attention and RMSNorm. A learned reconstruction prior enables per-region attribution for explainable predictions. Pretraining uses masked-token reconstruction of normalized geometry statistics and multi-resolution contrastive consistency. The 10.9M-parameter backbone is pretrained on 61,052 CAD meshes from Thingi10K, MFCAD, and Fusion360. On a held-out split of 2,983 meshes, Shape achieves reconstruction R2 = 0.729 and 98.1% top-1 retrieval under the Wang-Isola protocol, with near-zero reconstruction train/val gap (contrastive scores use a larger evaluation pool). A 2x2 ablation on loss type and target-space normalization shows per-dimension normalization is critical: without it, performance collapses (R2 < 0.14, top-1 < 88%); with it, both losses succeed (R2 > 0.70, top-1 > 96%). Smooth-L1 offers secondary stability. Code, embeddings, and an interactive demo are released at https://github.com/simd-ai/shape.

Recent text-to-scene generation approaches largely reduced the manual efforts required to create 3D scenes. However, their focus is either to generate a scene layout or to generate objects, and few generate both. The generated scene layout is often simple even with LLM's help. Moreover, the generated scene is often inconsistent with the text input that contains non-trivial descriptions of the shape, appearance, and spatial arrangement of the objects. We present a new paradigm of sequential text-to-scene generation and propose a novel generative model for interactive scene creation. At the core is a 3D Autoregressive Diffusion model 3D-ARD+, which unifies the autoregressive generation over a multimodal token sequence and diffusion generation of next-object 3D latents. To generate the next object, the model uses one autoregressive step to generate the coarse-grained 3D latents in the scene space, conditioned on both the current seen text instructions and already synthesized 3D scene. It then uses a second step to generate the 3D latents in the smaller object space, which can be decoded into fine-grained object geometry and appearance. We curate a large dataset of 230K indoor scenes with paired text instructions for training. We evaluate 7B 3D-ARD+, on challenging scenes, and showcase the model can generate and place objects following non-trivial spatial layout and semantics prescribed by the text instructions.

We introduce a hybrid Gaussian-hash-grid radiance representation for reconstructing 2D Gaussian scene models from multi-view images. Similar to NeST splatting, our approach reduces the entanglement between geometry and appearance common in NeRF-based models, but adds per-Gaussian latent features alongside hash-grid features to bias the optimizer toward a separation of low- and high-frequency scene components. This explicit frequency-based decomposition reduces the tendency of high-frequency texture to compensate for geometric errors. Encouraging Gaussians with hard opacity falloffs further strengthens the separation between geometry and appearance, improving both geometry reconstruction and rendering efficiency. Finally, probabilistic pruning combined with a sparsity-inducing BCE opacity loss allows redundant Gaussians to be turned off, yielding a minimal set of Gaussians sufficient to represent the scene. Using both synthetic and real-world datasets, we compare against the state of the art in Gaussian-based novel-view synthesis and demonstrate superior reconstruction fidelity with an order of magnitude fewer primitives.

Understanding and generating 3D objects as compositions of meaningful parts is fundamental to human perception and reasoning. However, most text-to-3D methods overlook the semantic and functional structure of parts. While recent part-aware approaches introduce decomposition, they remain largely geometry-focused, lacking semantic grounding and failing to model how parts align with textual descriptions or their inter-part relations. We propose DreamPartGen, a framework for semantically grounded, part-aware text-to-3D generation. DreamPartGen introduces Duplex Part Latents (DPLs) that jointly model each part's geometry and appearance, and Relational Semantic Latents (RSLs) that capture inter-part dependencies derived from language. A synchronized co-denoising process enforces mutual geometric and semantic consistency, enabling coherent, interpretable, and text-aligned 3D synthesis. Across multiple benchmarks, DreamPartGen delivers state-of-the-art performance in geometric fidelity and text-shape alignment.

Diffusion models have seen widespread adoption for text-driven human motion generation and related tasks due to their impressive generative capabilities and flexibility. However, current motion diffusion models face two major limitations: a representational gap caused by pre-trained text encoders that lack motion-specific information, and error propagation during the iterative denoising process. This paper introduces Reconstruction-Anchored Diffusion Model (RAM) to address these challenges. First, RAM leverages a motion latent space as intermediate supervision for text-to-motion generation. To this end, RAM co-trains a motion reconstruction branch with two key objective functions: self-regularization to enhance the discrimination of the motion space and motion-centric latent alignment to enable accurate mapping from text to the motion latent space. Second, we propose Reconstructive Error Guidance (REG), a testing-stage guidance mechanism that exploits the motion diffusion model's inherent self-correction ability to mitigate error propagation. At each denoising step, REG uses the motion reconstruction branch to reconstruct the previous estimate, reproducing the prior error patterns. By amplifying the residual between the current prediction and the reconstructed estimate, REG highlights the improvements in the current prediction. Extensive experiments demonstrate that RAM achieves significant improvements and state-of-the-art performance. Our code will be released.

Melanoma is a highly aggressive skin cancer, making early and accurate diagnosis critical. While deep learning excels in skin lesion classification, standard ``black-box" models struggle to explain diagnostic uncertainty, limiting clinical trust. This work introduces a hybrid framework combining a class-aware adversarial Variational Autoencoder and an XGBoost classifier, transcending simple binary classification by leveraging a generative latent space for interpretable decision support. Guided by adversarial training, the model learns the visual characteristics of skin lesions and projects them into a continuous latent space, ensuring that similar images are grouped closely together. Trained on this latent space, the XGBoost classifier achieves a robust AUC of 0.868, competing closely with state-of-the-art models. For borderline cases, the framework enables clinicians to leverage the latent topology through Content-Based Image Retrieval. This provides a dual benefit: it allows the clinician to visually compare an ambiguous lesion against biopsy-confirmed precedents and acts as an early warning sign since a borderline classification can indicate that a lesion shares features of both nevi and melanomas, potentially requiring close monitoring. Our approach translates algorithmic hesitation into transparent, evidence-based visual support, bridging the gap between predictive performance and clinical trust.

Multimodal large language models (MLLM) for surgical scene understanding typically inject hundreds of dense visual tokens into a language model, leading to costly inference and limited spatial traceability for generated answers. We present Slot2Text, a dual-mode surgical MLLM that replaces dense representations of visual input with a compact set of regions encoded as slot latents. Instead of relying on contrastive alignment of the visual encoder with language, Slot2Text groups self-supervised vision features into a few regions--slots that are consumed by the language model as area-labeled visual tokens. Slot2Text-Fast uses the slot prefix to answer surgical questions. Slot2Text-Reason also identifies and locates areas relevant for reasoning, linking language outputs to corresponding slot tokens, masks or regions. Experiments on multiple visual question answering and visual grounding benchmarks show that Slot2Text-Fast is competitive with state-of-the-art baseline at a much lower cost, reducing the average total token consumption by a 91.8% and the visual prefix from 1,295 to 47 tokens (a 96.4% reduction). Slot2Text-Reason trades additional tokens and latency for explicit area identities, locations, and traceable spatial evidence. These results establish compact slot latents as an efficient default visual interface for surgical MLLMs, with grounded reasoning invoked when greater spatial traceability is required.