3D Encoder

Wireless foundation models (WFMs) have emerged as a promising paradigm for unified channel state information (CSI) acquisition across diverse tasks in sixth-generation (6G) networks. Although WFMs significantly outperform task-specific small models, their zero-shot cross-scenario generalization still remains limited for real-world applications. Existing positional embeddings, the sole interface through which self-attention perceives the temporal-frequency-antenna 3D physical coordinates of CSI, fail to capture the highly dynamic and axis-dependent coherence inherent in wireless channels. This paper proposes Adaptive 3D-RoPE, a channel-driven 3D rotary positional embedding framework for WFMs to dynamically align the 3D positional embeddings with the instantaneous coherence state of heterogeneous CSI. The design proceeds in three stages: first, an axis-wise learnable rotary prior independently preserves the temporal, frequency, and antenna coordinate structures; second, a feature-guided rotary modulation module maps the feature-wise standard deviation of visible CSI tokens to compact, sample-adaptive scales; third, identical coordinate offsets induce dynamically adjusted query-key interactions tailored to the instantaneous channel state. Extensive experiments on both simulated and measured datasets validate the effectiveness of Adaptive 3D-RoPE across three complementary dimensions. It reduces NMSE by 10.14, 6.25, and 4.61 dB relative to baselines under antenna, temporal, and frequency scaling, respectively. It transfers effectively to real-world measured CSI and remains robust under imperfect CSI. Finally, it transfers to the independently designed LWM backbone and beam-prediction task, improving zero-shot Top-1 accuracy by 8.03 percentage points.

Modern visual world modeling systems increasingly rely on high-capacity architectures and large-scale data to produce plausible motion, yet they often fail to preserve underlying 3D geometry or physically consistent camera dynamics. A key limitation lies not only in model capacity, but in the latent representations used to encode geometric structure. We propose S$^2$VAE, a geometry-first latent learning framework that focuses on compressing and representing the latent 3D state of a scene, including camera motion, depth, and point-level structure, rather than modeling appearance alone. Building on representations from a Visual Geometry Grounded Transformer (VGGT), we introduce a novel type of variational autoencoder using a product of Power Spherical latent distributions, explicitly enforcing hyperspherical structure in the bottleneck to preserve directional and geometric semantics under strong compression. Across depth estimation, camera pose recovery, and point cloud reconstruction, we show that geometry-aligned hyperspherical latents consistently outperform conventional Gaussian bottlenecks, particularly in high-compression regimes. Our results highlight latent geometry as a first-class design choice for physically grounded visual and world models.

Implicit neural representations (INRs) are increasingly being used as tools to map coordinates to signals, encompassing applications from neural fields to texture compression, shape representations, and beyond. Most INR methods are based on using high-dimensional projections of the initial coordinates through encoders such as grid or positional encoding. Nevertheless, positional encoding is often insufficient and grids, as we show in this paper, require high resolution for being able to learn. In this paper, we demonstrate that positional encoding can be used not only as a high-dimensional embedding but also decomposed as a series of meaningful points. We propose the Positional Encoding Projected Sampling, where we treat the projection of the original coordinate at each frequency as a point of interest. We describe the motion of each point with respect to the frequencies and show that it follows a unique pattern. Finally, we use the unique motion of each point as a basis decomposition for doing learned positional encoding using grids. We prove, using three competitive applications; image representation, texture compression, and signed distance function; that the proposed approach outperforms the current state of the art methods, and often requires 25% less parameters for equivalent reconstruction error or rendering.

Single-view 3D shape retrieval is a fundamental yet challenging task that is increasingly important with the growth of available 3D data. Existing approaches largely fall into two categories: those using contrastive learning to map point cloud features into existing vision-language spaces and those that learn a common embedding space for 2D images and 3D shapes. However, these feed-forward, holistic alignments are often difficult to interpret, which in turn limits their robustness and generalization to real-world applications. To address this problem, we propose Pose-Aware 3D Shape Retrieval (PASR), a framework that formulates retrieval as a feature-level analysis-by-synthesis problem by distilling knowledge from a 2D foundation model (DINOv3) into a 3D encoder. By aligning pose-conditioned 3D projections with 2D feature maps, our method bridges the gap between real-world images and synthetic meshes. During inference, PASR performs a test-time optimization via analysis-by-synthesis, jointly searching for the shape and pose that best reconstruct the patch-level feature map of the input image. This synthesis-based optimization is inherently robust to partial occlusion and sensitive to fine-grained geometric details. PASR substantially outperforms existing methods on both clean and occluded 3D shape retrieval datasets by a wide margin. Additionally, PASR demonstrates strong multi-task capabilities, achieving robust shape retrieval, competitive pose estimation, and accurate category classification within a single framework.

Vector Symbolic Architectures (VSAs) provide a well-defined algebraic framework for compositional representations in hyperdimensional spaces. We introduce HyperSpace, an open-source framework that decomposes VSA systems into modular operators for encoding, binding, bundling, similarity, cleanup, and regression. Using HyperSpace, we analyze and benchmark two representative VSA backends: Holographic Reduced Representations (HRR) and Fourier Holographic Reduced Representations (FHRR). Although FHRR provides lower theoretical complexity for individual operations, HyperSpaces modularity reveals that similarity and cleanup dominate runtime in spatial domains. As a result, HRR and FHRR exhibit comparable end-to-end performance. Differences in memory footprint introduce additional deployment trade-offs where HRR requires approximately half the memory of FHRR vectors. By enabling modular, system-level evaluation, HyperSpace reveals practical trade-offs in VSA pipelines that are not apparent from theoretical or operator-level comparisons alone.

Flow-matching methods for 3D shape assembly learn point-wise velocity fields that transport parts toward assembled configurations, yet they receive no explicit guidance about which cross-part interactions should drive the motion. We introduce TORA, a topology-first representation alignment framework that distills relational structure from a frozen pretrained 3D encoder into the flow-matching backbone during training. We first realize this via simple instantiation, token-wise cosine matching, which injects the learned geometric descriptors from the teacher representation. We then extend to employ a Centered Kernel Alignment (CKA) loss to match the similarity structure between student and teacher representations for enhanced topological alignment. Through systematic probing of diverse 3D encoders, we show that geometry- and contact-centric teacher properties, not semantic classification ability, govern alignment effectiveness, and that alignment is most beneficial at later transformer layers where spatial structure naturally emerges. TORA introduces zero inference overhead while yielding two consistent benefits: faster convergence (up to 6.9$\times$) and improved accuracy in-distribution, along with greater robustness under domain shift. Experiments on five benchmarks spanning geometric, semantic, and inter-object assembly demonstrate state-of-the-art performance, with particularly pronounced gains in zero-shot transfer to unseen real-world and synthetic datasets. Project page: https://nahyuklee.github.io/tora.

Deformable 3D Gaussian Splatting (3DGS) has emerged as a popular method for real-time talking head synthesis, offering high-quality renderings at low latency. Tri-planes are a common choice for encoding Gaussians prior to deformation since they provide a compact and continuous representation. However, tri-plane encodings are limited by grid resolution and approximation errors introduced by projecting 3D volumetric fields onto 2D subspaces. Recent work has demonstrated the effectiveness of per-Gaussian embeddings for driving temporal deformations in 4D scene reconstruction. We introduce EmbedTalk, which leverages these embeddings to model speech-driven facial deformations for talking head synthesis. Comprehensive experiments show that EmbedTalk improves rendering quality, lip synchronisation, and motion consistency over previous 3DGS-based methods, while remaining competitive with state-of-the-art generative models. Replacing tri-plane features with embeddings also yields significantly more compact models that achieve 60+ FPS on a laptop GPU (RTX 2060 6 GB). Our code will be placed in the public domain on acceptance.

In real scenarios, videos can span several minutes or even hours. However, existing research on spatio-temporal video grounding (STVG), given a textual query, mainly focuses on localizing targets in short videos of tens of seconds, typically less than one minute, which limits real-world applications. In this paper, we explore Long-Form STVG (LF-STVG), which aims to locate targets in long-term videos. Compared with short videos, long-term videos contain much longer temporal spans and more irrelevant information, making it difficult for existing STVG methods that process all frames at once. To address this challenge, we propose an AutoRegressive Transformer architecture for LF-STVG, termed ART-STVG. Unlike conventional STVG methods that require the entire video sequence to make predictions at once, ART-STVG treats the video as streaming input and processes frames sequentially, enabling efficient handling of long videos. To model spatio-temporal context, we design spatial and temporal memory banks and apply them to the decoders. Since memories from different moments are not always relevant to the current frame, we introduce simple yet effective memory selection strategies to provide more relevant information to the decoders, significantly improving performance. Furthermore, instead of parallel spatial and temporal localization, we propose a cascaded spatio-temporal design that connects the spatial decoder to the temporal decoder, allowing fine-grained spatial cues to assist complex temporal localization in long videos. Experiments on newly extended LF-STVG datasets show that ART-STVG significantly outperforms state-of-the-art methods, while achieving competitive performance on conventional short-form STVG. Our code is at: https://github.com/HengLan/ART-STVG.

Robot manipulation uses temporal context to select actions and visual foresight to assess their consequences, yet dense representations of past and future observations incur substantial processing costs. We introduce PACT-WAM, a world-action model that jointly generates a 16-step action trajectory and its temporally corresponding visual forecast through conditional flow sampling. Hierarchical history encoding assigns coarse spatial representations to earlier observations and finer representations to recent ones, retaining 16 observations with 256 tokens per view, 75% fewer than dense encoding of the same frames. A shared flow module jointly updates continuous action and visual states through two modality-specific heads under transition-wise causal attention, and a TiTok-VAE decoder reconstructs multi-view future images from the visual latents. Decoded forecasts also support Proposal Review (PR), a vision-language model component for execution-prefix selection and proposal rejection. Without PR, PACT-WAM achieves average success rates of 98.6%, 92.3%, and 78.0% on LIBERO, RoboTwin 2.0, and real-world Piper tasks, respectively. PR provides a test-time enhancement, raising these rates to 99.5%, 93.4%, and 86.7%. Ablations show that hierarchical history allocation and joint action-visual generation improve control success, while analyses of visual capacity and forecast-guided execution characterize the trade-offs between success and proposal-generation cost.

The zero-shot evaluation of time series foundation models (TSFMs) for classification typically uses a frozen encoder followed by a task-specific classifier. However, this practice violates the training-free premise of zero-shot deployment and introduces evaluation bias due to classifier-dependent training choices. To address this issue, we propose TIC-FM, an in-context learning framework that treats the labeled training set as context and predicts labels for all test instances in a single forward pass, without parameter updates. TIC-FM pairs a time series encoder and a lightweight projection adapter with a split-masked latent memory Transformer. We further provide theoretical justification that in-context inference can subsume trained classifiers and can emulate gradient-based classifier training within a single forward pass. Experiments on 128 UCR datasets show strong accuracy, with consistent gains in the extreme low-label situation, highlighting training-free transfer for time series classification.The source code is publicly available at https://github.com/fangjuntao/TIC-FM.

Reconstructing natural images from fMRI requires bridging neural activity with both the structural and semantic representations used by modern generative models. Existing diffusion-based decoders often condition on a single global fMRI embedding, which limits their ability to exploit the hierarchical organization of the visual cortex and makes the contribution of different visual areas difficult to inspect. We propose Hi-DREAM, a brain-inspired hierarchical diffusion framework that structures fMRI conditioning according to early, middle, and late visual Regions of Interest (ROI) streams. A ROI adapter converts these streams into a multi-scale cortical pyramid, and a lightweight ROI-conditioned ControlNet injects the resulting anatomy-aware priors into matched U-Net depths during denoising. Experiments on the Natural Scenes Dataset (NSD) show that Hi-DREAM achieves state-of-the-art high-level semantic reconstruction while retaining strong low-level structure. Further ablation and attribution analyses show that the proposed hierarchy-aware conditioning is effective, and that different ROI streams provide complementary, inspectable contributions to reconstruction.

We study latent geometry as an explicit component of representation quality in data-scarce learning. For an encoder (φ), we define (Q_{β,γ}(φ)=I(φ(X);Y)-β\mathcal C(φ)-γd_{\mathrm{int}}(φ)), combining task-relevant information with penalties for curvature and intrinsic latent dimension. Thus geometry becomes part of the bottleneck criterion, not only a post hoc diagnostic. Under smooth-manifold, loss-transfer, and estimator-concentration assumptions, we derive non-asymptotic low-label generalization bounds where intrinsic dimension and covering complexity enter explicitly. We characterize the information--geometry frontier and prove empirical-surrogate consistency. The analysis links encoder geometry to learning through latent covering numbers, loss-class entropy, and uniform deviation. We instantiate the theory as \texttt{V-GIB}, adding curvature and dimension penalties to variational bottleneck training. Real low-label benchmarks compare \texttt{V-GIB} with ERM, VIB, and ablations across (1%)--(20%) label fractions. Results show improved performance and reduced geometric complexity in several regimes, especially FashionMNIST and CIFAR-10, while confirming that no fixed regularizer is universally dominant.

This paper presents DINO-SLAM, a DINO-informed design strategy to enhance implicit (Neural Radiance Field -- NeRF) and explicit representations (Gaussian Splatting -- GS) in SLAM systems through the more comprehensive semantics understanding enabled by DINO. This latter alone, however, lacks proper 3D geometry understanding, allowing only for marginal improvements. Therefore, we rely on a Scene Geometry Encoder (SGE) to enrich DINO features into geometry-aware DINO features (geoDINO), to better understand those geometric relationships that vanilla DINO features fail to capture. Building upon it, we propose two foundational paradigms for NeRF and GS SLAM systems integrating geoDINO features. Compared to state-of-the-art methods, our DINO-informed pipelines achieve superior performance on the Replica, ScanNet, and TUM datasets.

Traffic prediction is difficult due to the complex interplay of temporal evolution, spatial interactions, and delayed spatio-temporal propagation over road networks. Existing methods either model spatial and temporal dependencies separately or employ unified spatio-temporal structures, but they often insufficiently characterize how neighboring sensors at historical timestamps influence a target node, while complex joint models may incur high computation. This paper proposes STEI-PCN, an efficient pure convolutional network based on spatio-temporal encoding and relation inference. It first builds a local causal joint spatio-temporal graph to restrict candidate interactions, then uses absolute position and relative distance encodings to infer dynamic edge weights. A single-layer graph convolution with a position-aware gated activation unit captures local joint dependencies, and temporal dilated causal convolutions complement long-range temporal patterns. A multi-view prediction module fuses raw, local propagation, and long-range temporal representations for direct multi-step forecasting. Experiments on PeMS03, PeMS04, PeMS07, PeMS08, and PeMS-Bay under multiple horizons show that STEI-PCN achieves competitive accuracy with moderate parameters and low training and inference costs. Ablation and fluctuation analyses further verify the contributions of the main components and empirically analyze the effects of the training-stage constraints under sharp speed changes. Our code is available at a GitHub link https://github.com/Jessez2/STEI-PCN.

Generative medical event models use tokenized sequences of patient timelines as input, but practical guidance on the many decisions around tokenization is limited. We benchmark quantization granularity, reference-range anchoring, code--value fusion, numeric and temporal encodings, and native versus harmonized event representations from an expert-mapped common data model. Using both Llama and Qwen architectures, 156 models were trained from three initialization seeds, with each configuration following a shared training recipe for up to five epochs. We evaluated learned representations from the first 24 hours of hospitalization with linear probes to predict binary and continuous outcomes during hours 24-48. Fused tokens pairing codes with value deciles increased performance across all eight outcome families relative to the equivalent unfused tokenized input with area under the receiver operating characteristic curve (AUROC) gains of $+0.002$ to $+0.033$ and Spearman correlation gains of $+0.025$ to $+0.114$. Neither anchoring value bins to reference ranges nor increasing quantization granularity consistently improved performance, while xVal variants underperformed both discrete and soft encodings. Alternatives to explicit time tokens, such as event-order and admission-relative rotary position embeddings (RoPE), yielded higher family-mean point estimates across all eight families while reducing input length. When evaluating native input against input mapped to the Common Longitudinal Intensive Care Unit Data Format (CLIF), the CLIF pipeline used 30.8% as many training tokens while improving performance in six of eight outcome families. These findings show tokenization and event encoding are consequential design choices when learning patient representations for downstream classification and regression.