Robotic world models predict future video frames from the current observation and a sequence of robot actions. While recent models follow the given actions closely, manipulated objects often collapse or move inconsistently with the action. We attribute this to the conditioning: an action, or a rendering of the robot, specifies where the robot goes but not how the scene responds, leaving a single video model to infer the interaction dynamics implicitly while also rendering appearance. We therefore separate dynamics prediction from visual rendering, with 3D point dynamics as the proxy between them. Specifically, we lift the first frame into 3D points, simulate how they respond to the robot's motion with a learned point dynamics model, and render the simulated points back into pixel-aligned videos that any video model can read. In controlled experiments on the DROID dataset with two video renderers of different architectures, this point dynamics proxy reduces object motion error by 8.6% and 6.2% without degrading image quality, and by 8.1% on a real bimanual robot. Moreover, the video follows the simulation: more accurate dynamics yield more accurate object motion, and editing the points steers the video. This suggests that advances in dynamics modeling can translate into more faithful robotic world models.
Facial deepfake detectors are usually trained on blending fakes, where a generated face is pasted back into a real frame. Recent generators synthesize the entire face and leave no traces of blending. On a benchmark of 19 datasets released from 2019 to 2026, the best prior facial deepfake detector reaches a mean AUROC of 96.2% on ten blending datasets, but none exceeds 78.1% on five datasets of entire face synthesis (EFS). AI-generated content (AIGC) detectors do not close this gap. We propose TwinD, a detector trained on FaceForensics++ fakes and two kinds of twins: fakes made from a real frame. A reconstruction twin is a real frame reconstructed by an autoencoder, with the real pixels put back in a random region. It targets the neural traces of a decoder. A blob twin blends an altered copy of a real frame with the frame itself, targeting blending artifacts. It needs no generator and no face-shaped mask. Together, these twins provide two complementary forensic cues that generalize to unseen manipulations. TwinDreaches a mean AUROC of 96.3%, 8.9 points above the prior state-of-the-art, with 98.1% on the blending datasets and 95.6% on the EFS datasets. Trained on twins alone, without a deepfake dataset, single models still reach 94.3% on average over three seeds. For reproducibility, the code, trained model, and training data are at: https://github.com/yermandy/TwinD
3D Gaussian Splatting (3DGS) enables real-time novel view synthesis, but Gaussian rasterization attenuates fine-scale scene structure through the projected covariance of each primitive. We show that contracting these projected covariances at inference time can mitigate the effective bandwidth limitation induced by finite Gaussian footprints, motivating a multi-aperture rendering framework for super-resolution. Based on this insight, we propose AGSR (Aperture-Guided Super-Resolution), a lightweight inference-time framework that progressively fuses multi-aperture renderings using uncertainty-guided feature aggregation and geometric conditioning from transmittance and depth cues. AGSR operates as a plug-in enhancement for pre-trained 3DGS models. Experiments on the Mip-NeRF 360 benchmark show that AGSR outperforms lightweight image-space super-resolution baselines, with gains increasing as the underlying Gaussian representation deteriorates. The results provide evidence that covariance contraction exposes bandwidth-sensitive structural cues attenuated by native Gaussian rasterization. AGSR achieves these improvements while maintaining real-time performance with up to 139k parameters.
Flat color-block drafts specify layout and regional color while leaving interior detail unspecified, yet pretrained flow-based generators can preserve their flat appearance even under photorealistic prompts. We trace this behavior to a cross-scale mechanism linking fine pixel variation to regional means in variational autoencoder (VAE) latent space. Flattening photo regions systematically shifts their latent means, with offsets concentrated along a shared channel direction within each backbone. Controlled field exchanges in FLUX.1-dev and Qwen-Image establish the effect of this displacement on texture generation: adding the photo-derived regional mean field to a flat latent restores texture, whereas adding its complementary field alone largely preserves flatness. A second-order encoding expansion explains how encoder curvature can convert zero-mean perturbations into mean shifts. An analytically solvable conditional Gaussian flow shows how regional noise averaging enables near-unit endpoint sensitivity at high signal-to-noise ratios. This mechanism motivates training-free rendering through pixel dither before encoding, mean-field scaling, and early boundary holding. Paired low-rank adaptation (LoRA) with draft conditioning improves content recovery and produces photo-like regional means under the standard velocity objective. Explicit regional mean guidance remains effective after adaptation, continuously controlling texture and improving color fidelity. These findings connect encoding shifts to persistent generation behavior and support interventions evaluated for texture, spatial fidelity, and semantic recovery.
Vision Language Models (VLMs) have shown remarkable capabilities in qualitative spatial reasoning, but remain limited in quantitative spatial reasoning. Current work leverages the VLM agent to invoke external vision tools to perform computation. However, these methods usually bind each functional requirement to a single predefined vision specialist, implicitly resulting in fixed and rigid tool pipelines. In practice, addressing diverse spatial scenarios and tasks demands a broader diversity in pipeline orchestration and vision specialist selection to achieve robust spatial reasoning. However, such diversity leads to a combinatorial execution space, making it challenging for the agent to orchestrate the optimal pipelines and specialists for different scenarios. In this work, we propose SpatialMem, a memory framework that enables a VLM agent to accumulate tool-use experience and leverage that memory to orchestrate vision specialists to solve the spatial reasoning problem. SpatialMem builds a two-level memory structure, with task memory capturing tool orchestration and scene memory capturing tool selection. It further introduces a dual-state trajectory optimization process to preserve successful tool experiences and refine suboptimal ones. Moreover, existing spatial reasoning benchmarks suffer from quality issues and limited coverage of scene types and tasks. To address these gap, we propose MetricBench, a benchmark covering both indoor and outdoor scenes and three quantitative spatial reasoning tasks. Our method achieves state-of-the-art results on MetricBench, Q-Spatial, and SpatilRGPT-Bench, outperforming best VLM baselines by 32% on MetricBench. Additionally, the memory design yields a 9.1% absolute gain over a memory-free agent on MetricBench.
In this paper we describe a method for tracking rigid objects using one or several cameras. The tracking process consists of aligning a 3-D model representation of an object with image contours by measuring and minimizing the image error between predicted model points and image contours. The tracker behaves like a visual servo loop where the internal and external camera parameters are updated at each new image acquisition. We study in detail the Jacobian matrix associated with this minimization process in the presence of both point-to-point and point-to-contour matches. We establish the minimal number of matches that are needed as well as the singular configurations leading to a rank-defficient Jacobian matrix. We find a mathematical link between the point-to-point and point-to-contour cases. Based on this link we show that the latter has the same kind of singularities than the former. Moreover, we study multiple camera configurations which optimize the robustness of the method in the presence of single-camera singularities, bad, noisy, or missing data. Extensive experiments done with a complex ship part and with up to three cameras validate the method. In particular we show that the tracker may well be used as a camera calibration procedure.
Self-supervised foundation models are often said to need fewer labels than supervised training. We test this on a forestry task: classifying 20 tree species classes in 0.2 m aerial images from the TreeSatAI dataset. We compare a frozen DINOv3 model pretrained on 493 million satellite images, with only a linear classifier trained on top, against an ImageNet ResNet-50 that is fine-tuned end to end. On the official split, fine-tuning wins clearly (macro F1 0.764 vs. 0.685). However, 28.8% of the official test patches overlap a training or validation patch, and copying the label of the nearest such patch is correct for 83.6% of them. We therefore build a spatial split that holds out whole 2 km squares with a 200 m buffer. On this split the comparison becomes inconclusive (0.648 vs. 0.640, paired 95% CI of the difference [-0.054, 0.016]), and the learning curves of the two models overlap at every label budget from 343 to 33,126 training patches. On the five rarest classes, the frozen model achieves a higher mean F1 (0.566 vs. 0.445), although the confidence interval for this difference includes zero ([-0.05, 0.23]). Diagnostics show that the frozen features mainly miss differences between closely related species, such as the two native oaks. Choosing the input size and layer on validation data (224 px, block 16) raises the frozen model to 0.662 on the spatial split. A comparison that looks decisive on the official split is thus inconclusive under geographically separated evaluation. We release the spatial split and our code.
Although memory-augmented architectures have expanded the ability of large vision-language models (LVLMs) to process long videos, existing retrieval processes still largely rely on global semantic matching and lack state-driven temporal control, often returning fragmented or redundant evidence and failing to maintain a coherent evidence chain. To address this challenge, we propose EvoMem, a time-aware long-video memory agent guided by self-evolving retrieval skills. EvoMem constructs a hierarchical time-indexed memory, where different types of evidence are aligned onto a unified video timeline; during retrieval, temporal cues from acquired evidence are further converted into explicit signals that constrain subsequent memory access. To address recurring procedural errors in long-video question answering, such as unclear retrieval targets, improper memory access paths, or premature stopping, EvoMem introduces retrieval skills to provide process-level guidance, enabling the agent to dynamically adjust retrieval content, memory operations, and target temporal ranges according to the current evidence state, while deciding when to stop. Furthermore, inspired by reinforcement learning, we formulate skill evolution as a reward-guided policy improvement process. Each candidate skill is treated as an external policy patch, and rewards are assigned according to answer-correctness state transitions on the validation set. In this way, effective skills are accepted and low-quality skills are rejected without updating model parameters. Experiments on four long-video question answering benchmarks show that EvoMem significantly outperforms existing baselines, achieving an average accuracy of 60.1% and an average improvement of 8.6 percentage points over the previous best methods, demonstrating its effectiveness for long-video reasoning.
Accurately extracting and mapping road surface elements is crucial for applications such as high-definition maps, road asset management, and intelligent transportation systems. This study first provides a comprehensive review of the benchmark datasets for extracting and mapping road elements, encompassing overhead-to-perspective data, large-to-local scales, and coarse-to-fine details. Subsequently, we thoroughly review the relevant approaches for extracting road surface elements. However, we argue that existing datasets still require enhancements in semantic richness, data modality diversity, and scenario complexity. To bridge these gaps, this paper presents WHU-Road3D, a large-scale multimodal dataset for fine-grained road surface mapping to address these shortcomings. WHU-Road3D provides approximately 100 km of diverse Chinese highways and urban streets. It features a rich combination of MLS point clouds, associated panoramic images, vehicle trajectories, detailed vectorized annotations, and pointwise instance annotations for road surface elements. The dataset carefully annotates 15 distinct categories of road elements, including common markings like road boundaries, zebra crossings, lane dividers, stop lines, as well as 10 types of direction symbols. To our knowledge, WHU-Road3D stands as the first multimodal dataset for mapping such a comprehensive range of road surface elements, designed to facilitate cutting-edge research in instance segmentation, full-element extraction and mapping on road surfaces. Beyond these immediate applications, WHU-Road3D also provides a valuable source for hierarchical road topology reasoning, map updating, and street scene generation. The dataset, along with the code for our benchmark methods, will be made publicly available at \href{https://github.com/WHU-USI3DV/WHU-Road3D.git}{https://github.com/WHU-USI3DV/WHU-Road3D.git}.
Recognizing when a requested spatial condition holds in streaming video can require maintaining and updating spatial knowledge across changing viewpoints. We study this capability through proactive spatial reasoning, where a model receives a query in advance and determines whether and when to respond as the stream unfolds. We introduce ProSpace-Bench, comprising nine tasks across object, camera, and scene reasoning, and ProSpace-Train for supervised training. To support spatial decisions, we propose ProSpace-VLM, which injects explicit geometric encodings estimated causally from incoming frames into intermediate VLM layers, using the existing KV cache as spatial memory and spatial token selection for efficient streaming. Experiments reveal a substantial gap between VLMs and humans, while ProSpace-VLM improves timely responses and correct abstention over fine-tuned baselines. Analyses examine how ProSpace-VLM represents previously observed spatial information across changing viewpoints.
Recalling where and when an object was last seen and who saw it requires recognizing it across days and people. Existing long-video benchmarks focus on events, object categories, or relative locations, leaving instance-level spatial memory largely untested in multimodal models. To address this gap, we present EgoWitness, a benchmark for instance-level spatial memory and reasoning in egocentric video. Its 8,453 human-verified questions span approximately 324 hours of week-long and short-horizon recordings and test seven abilities, from recognizing an object or room to recalling who saw it and reasoning about room visits. Over one million reviewed masks of 34 registered landmarks and 21 movable-object labels are linked through room timelines. Among twelve direct vision-language models (VLMs), open-weight models reach at most 46.97% accuracy, and none exceeds 35% on identifying who saw an object last. We therefore propose EgoClue, a spatial-memory agent that retrieves candidate intervals, verifies instance identity, and computes spatial and temporal relations with five tools. Two-stage training with supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) raises its 4B controller from 23.64% to 72.65% accuracy on 1,320 held-out questions. Evidence-trace synthesis matches the best direct VLM at 74.85% on week-long questions and achieves 80.56% in environments excluded from training.
Data-driven weather forecasting requires distinguishing local states and their temporal evolution. However, learned patterns remain implicit in neural forecasters, making their influence difficult to interpret. Failing to distinguish local evolution can lead to similar forecasts for locations that develop differently, increasing prediction error. This motivates explicit references for distinguishing and interpreting local patterns. Prototypes provide explicit pattern references, but sample- or task-level conditioning leaves open how to adapt to distinct locations and variable groups. To deal with this problem, we design ProtoCast, which uses group-specific prototype banks to adapt local features while accounting for cross-variable dependencies. Specifically, (1) we combine each group's historical features and temporal change with forecast context from cross-variable attention for evolution-aware prototype matching. (2) Sparse mixtures of prototype-specific low-rank operators tailor corrections to each group and location. A gated attention block then exchanges the corrected features across groups before prediction. Each group's correction decomposes into explicit prototype contributions before this fusion. (3) We jointly train the forecasting network and operators from scratch and refit each group's prototypes before every training epoch. Slot tracking preserves prototype-operator associations, making each prototype's contribution to local corrections more interpretable. Experiments on GlobalForecast and TyphoonTracking show improved forecasting accuracy, while prototype deletion links local feature contributions to changes in prediction error. This work advances the development of more accurate and interpretable data-driven weather forecasting models. Project page: https://zhixuanli.github.io/projects/li2026protocast
Flow matching enables high-quality image and video generation, but its iterative sampling process requires repeated evaluations of large neural networks. Existing few-step methods can distort the output and reduce its agreement with a many-step reference. We propose TrajCal, a training-free calibration framework for few-step sampling. Based on scatter search, a well-established metaheuristic for difficult optimization problems, TrajCal starts with diverse seed schedules, deforms and combines these candidates, and selects one fixed schedule by final output fidelity for reuse on new inputs. For both image and video generation, TrajCal achieves higher reference fidelity than all compared baselines at low evaluation budgets while maintaining competitive visual quality. The search is effective even with randomly generated seeds, while trajectory-based seeds further improve fidelity. For cross-style generalization, photography-calibrated schedules also improve fidelity on unseen image styles.
Medical vision-language agents increasingly rely on segmentation tools, and recent work has focused on acquiring the right mask, often treating a good mask as the end point of tool use. But do these agents use the resulting evidence effectively? To answer this question, we need a benchmark that separates evidence use from segmentation quality, spans multiple medical image types, and allows the same mask to be presented in different ways. Existing medical VQA datasets do not jointly provide these properties. We therefore construct Mask-Evidence MedVQA, a 2.5K-example benchmark spanning four image types, where each question is paired with an answer-relevant ground-truth mask and multiple possible renderings. Our study reveals two limitations of direct tool use. First, evidence presentation matters. Different renderings of the same fixed mask can substantially change model behavior. Second, masks introduce an evidence trade-off. Highlighting a relevant region can expose diagnostic evidence the model previously missed, but the added visual cue can also induce incorrect associations. As a result, even a ground-truth mask can correct some answers while turning initially correct ones into incorrect ones. Together, these findings show that acquiring the right mask is not enough. An agent must also decide how to present the evidence and whether it should change the answer. We therefore introduce TUNE (Tool Use is Not Enough), an evidence-management harness that explicitly manages what evidence to acquire, how to present it, and whether to accept the resulting answer update. Across three public medical VQA benchmarks, evidence management improves the fine-tuned Qwen3.5-9B backbone's mean accuracy by 4.0 percentage points. Full TUNE improves over direct tool use by 9.1 points on average across seven benchmarks and achieves competitive results against existing medical VLMs and agents.
Interactive game world models separate executable world evolution from visual generation, using video models to render engine-provided scene representations into detailed observations. During sustained gameplay, however, existing generative renderers struggle to follow changing conditions while preserving visual continuity across diverse proxies and target styles. We therefore introduce ContextRenderer, an autoregressive video diffusion framework designed for both dynamic conditioning and cross-proxy generalization. It couples Dynamic-Context Learning (DCL) for chunk-level interaction supervision with Recurrent History Compression (RHC) to follow condition updates while preserving visual continuity within a bounded token budget. Across proxy representations, Cross-Representation Flow Alignment (CRFA) connects proxy reconstruction with appearance and interaction variation through a shared high-fidelity reconstruction. WorldPrism, our interaction-centric dataset constructed from agent-built worlds, AAA gameplay, and open-source videos, supplies the corresponding renderings and chunk-level annotations for these learning tasks. Experiments on GWM-Bench and Solar-Bench show improved reconstruction and visual quality, while qualitative results demonstrate minute-long interactive rendering under changing prompts and subject references. Project page: https://rainzor.github.io/contextrenderer/
Spurious correlations can give image predictors high in-distribution accuracy but poor generalization under distribution shift. We introduce $\textbf{Rank2Replace}$, which transfers knowledge of task-relevant evidence from pretrained vision--language models (VLMs) to small predictors through training-time augmentation for out-of-distribution (OOD) generalization. The VLM ranks object-centric slots by task relevance and guides rank-dependent replacement with same-rank slot embeddings from another image. This augmentation retains the original image's label, encouraging reliance on preserved task evidence as incidental cues change, and supports combination with existing OOD methods. Across four datasets and six splits under shared frozen slot embeddings, the default cross-entropy configuration reaches macro-averaged average-group and worst-group accuracies of $80.43\%$ and $65.28\%$ without environment labels or dataset-specific replacement tuning. Slot corruption improves both macro-averaged metrics for all fourteen OOD training methods. Combining it with slot ranking and IRM reaches $81.96\%$ and $72.68\%$. Controlled ablations support the ranking and replacement design, with gains across five ranking VLMs and six extractor--predictor combinations. Permutation-invariant predictors further improve both macro-averaged metrics over their unaugmented counterparts while using VLM guidance only during training for efficient inference.
Precise forest species mapping using spaceborne hyperspectral data is challenging due to high spectral similarity and the inherent sub-pixel heterogeneity of natural forest stands. While traditional deep learning models exploit spatial context through convolutional architectures, these approaches can introduce noise in fragmented or mixed-species forests. In this paper, we propose a hierarchy-aware, multimodal deep learning framework to resolve sub-pixel forest composition by integrating PRISMA (PRecursore IperSpettrale della Missione Applicativa) hyperspectral imagery, high-resolution panchromatic data, and digital elevation models. The architecture features a multi-branch encoder combining an EfficientNetV2-based one-dimensional convolutional neural network (CNN) for spectral signatures and twin two-dimensional CNNs for spatial and topographic features, coupled with a novel structurally-masked Transformer. By using a binary transition tensor to condition the Transformer slot embeddings and constrain inference, the model enforces taxonomic consistency between broad forest categories and fine-grained species distributions. The framework is validated across three diverse Italian biogeographical regions (the Alps and the Central and Southern Apennines) using field-surveyed plots. The proposed method consistently outperforms state-of-the-art multi-label and ranking-based paradigms, particularly in identifying subordinate species and their relative dominance. The bidirectional learning mechanism allows sub-pixel compositional features to refine macroscopic forest type predictions, while the integration of topographic priors effectively resolves spectral ambiguities in complex terrains. This knowledge-driven method supports regional-scale biodiversity monitoring and forest characterization within the evaluated PRISMA scenarios.
We examine how well subspaces constructed from local curvature represent the change in the gradient update field after a finite step in semantic segmentation models. The local response is the action of the curvature operator on the actual input displacement. In a study of 50 images and two models, we measure the ratio of the loss increase attained in a fixed subspace to that attained in the full space. This ratio decreases as the search radius increases. On ten images from the same sample, we compare the one-dimensional subspace spanned by the initial gradient with two rank-16 subspaces for both models. The two equal-rank subspaces are assessed at the same initial states, using the same actual steps and measured vectors. Adding generalised Gauss-Newton directions to the initial gradient reduces the mean relative error in representing the field change from 0.922 to 0.874 for the first model and from 0.819 to 0.763 for the second. At the same rank, including the response direction for the initial gradient nearly eliminates the error in representing the local response. The mean error in representing the field change does not decrease. A decomposition with fixed pixel weights shows that the finite-step remainder is comparable in norm to the field change and makes the largest diagonal contribution to the squared projection error of the field after the step. Evaluation of a search subspace should measure how accurately it represents the field change and the local response.
Industrial bin picking requires pose estimates that are accurate under clutter, partial occlusion, and the reflective, texture-less nature of manufactured parts. Most recent approaches to 6D pose estimation follow multi-stage pipelines. These achieve high accuracy, but incur memory and runtime footprints that do not cope with the operational constraints of many real-world industrial deployments. Holistic pipelines estimate poses in a single pass, with cost controlled through indiscriminate subsampling or lowered resolution. Continuing this line, we present SCCV-6D, a fully sparse multi-view RGB-D estimator. We reconstruct the 3D space at millimeter precision, prune background via a staged heatmap mechanism, and predict class, translation offset and rotation at highest resolution for an arbitrary number of objects simultaneously. Rotation is formulated as canonical coordinate voting. Points vote for their canonical object coordinates with an associated confidence, enabling closed-form pose recovery with a symmetry-aware weighted Procrustes solve. Along an in-cluster point recovery induced by offset votes, this removes time-consuming post-processing stages. We reuse the task-supervised attention priors from the heatmap formulation for a novel attention-gated distillation objective that reweights teacher-student feature alignment. On the IPD test split we report competitive 0.826 AP at 1.13 s per scene, significantly faster than the leading entries. On MV-YCB our approach reaches 96.1 ADD-S, and remains competitive even on the densely packed bins of ROBI.
Temporal action localization (TAL) detectors must identify actions and estimate their temporal boundaries, but localization remains a major source of detection errors. While understanding what determines boundary placement requires faithful attribution of temporal evidence, existing attribution methods have not sufficiently addressed the temporal structure of TAL or the need to tailor attribution to boundary localization. We introduce a structure-aware attribution framework that averages contributions across random evidence-restoration orders following the Shapley-Shubik principle. We make it tractable by using the detector's receptive field to limit which time steps are ordered individually, and by approximating each order with a soft gradient path. For boundary predictions, we further introduce cell-structured restoration orders and boundary-anchored references that selectively remove adjacent evidence. We also adapt faithfulness evaluation to assess the joint effects of contiguous temporal evidence and the direction of boundary displacement. Experiments on THUMOS14 and EPIC-KITCHENS-100 with TriDet and ActionFormer demonstrate improved attribution faithfulness over existing methods. Attribution analysis and controlled interventions further reveal that class-score changes do not reliably track boundary displacement, providing input-level evidence that classification relies on content less informative about boundary placement. Code and evaluation artifacts will be publicly released.