New Computer Vision (cs.CV) papers, Wednesday 21 October 2026
185 papers announced on arXiv. Get them ranked for you, free →
Robotic Video World Models: A Survey of Applications, Research Challenges, Future Directions
AI summaryThis survey reviews video world models in robotics, highlighting their applications in policy learning, simulation, and planning. It also discusses key challenges like physical hallucinations and compute overhead, outlining future research directions.
Half-Truths Break Similarity-Based Retrieval
AI summaryThis paper shows that CLIP-style models incorrectly assign higher similarity scores to captions containing plausible but false details. The authors propose Component-Supervised CLIP (CS-CLIP), which fine-tunes models on decomposed caption units and foils to significantly improve compositional understanding.
FRAGMENT: Factorized Graph Representations for Document Generation and Editing via Entity-Aware Transformations
Structured documents such as invoices, forms, reports, and scientific articles derive meaning from the interplay between spatial layout, textual content, and logical structure. Generative models operating at the pixel or token level often struggle to capture these dependencies…
EGGHead: Efficient Generalizable Gaussian Heads from a Single Image
AI summaryEGGHead introduces a semantic Gaussian allocation algorithm for single-image 3D head reconstruction that operates in a single forward pass. It achieves superior reconstruction quality, novel view synthesis, and real-time rendering speeds while using a fixed Gaussian budget.
Beckmann World Models: Direct Terminal-Map Learning for One-Step Action-Conditioned Prediction
AI summaryThe paper introduces Beckmann World Models, which use terminal-map learning to directly predict future video frames from noise and actions in a single network call. This approach significantly reduces prediction error and computational overhead compared to existing baselines on robotic manipulation benchmarks.
Uncertainty Gates the Gradient: Adaptive Updates for Generalized Few-Shot 3D Segmentation
AI summaryUGA gates gradient updates using episodic uncertainty measured by intra‑class dispersion, allowing the model to learn from reliable 3D point‑cloud episodes while suppressing updates from noisy ones; on ScanNet200/++ it outperforms baselines by +3.08 HM IoU in 1‑/5‑shot settings.
Memory Internalization via Test-Time Training for Long-Video Understanding
AI summaryMIT^3 adapts a language model to a long video by writing the video into compact, video‑specific parameters during test‑time training, using causal tracing to locate where to adapt and visual memory distillation to capture visual evidence. It outperforms prior video‑memory methods on LVBench, LongVideoBench and Video‑MME with far fewer synthetic samples.
CodecMotion-OAD: Efficient Online Action Detection with Compressed-Domain Motion
AI summaryThis paper introduces CodecMotion-OAD, which replaces optical flow with video codec motion vectors for online action detection using a two-stage distillation training approach. It nearly matches RGB-flow accuracy while reducing processing time over elevenfold on CPUs.
GTSM: Gaussian Splatting with Thermal State-Space Dynamics for Spatio-Temporal Thermal Field Reconstruction and Prediction
AI summaryGaussian Splatting with Thermal State-Space Dynamics (GTSM) reconstructs and predicts 4‑D thermal fields from 2‑D observations using a shared latent process and Gaussian local states, achieving higher accuracy and better temporal extrapolation than baselines, and enabling novel‑view thermal rendering.
Codec-Compatible Learned Compression for 3D Gaussian Splatting
AI summaryCCT-GS improves video-based 3D Gaussian Splatting compression by applying a learned transform and scene adaptation to spherical harmonics maps. It achieves an average BD-rate reduction of 51.30% over standard video codecs.
Controllable Multi-Object Insertion for Detection Data Augmentation
Controllable object insertion provides a targeted approach to augmenting object detection datasets that require category--count fidelity, annotation alignment, and source-content preservation beyond visual realism. Several approaches have been proposed to address this grand…
CaMeWM: A Causal Medical World Model for Generative Tumor Evolution and Treatment Planning
AI summaryCaMeWM predicts counterfactual tumor evolution and survival after TACE by generating intervention-conditioned follow-up worlds and mapping them to restricted mean survival time. It achieves ~0.62 SSIM on CT transition and 0.615 C‑index, estimating a 9.6‑week RMST gain for the frozen policy.
SuperClip Pyramid for Video Temporal Grounding
AI summarySCVTG builds a multi‑scale temporal hierarchy of semantically coherent “SuperClips” and aggregates them with a temporally adaptive kernel, guided by relative saliency losses, achieving state‑of‑the‑art results on five video grounding benchmarks.
Rate as a Feature, Not a Nuisance: Curriculum FiLM Training for Event-Based Graph Neural Networks
AI summaryThis paper introduces a curriculum training method that conditions event-based Graph Neural Networks on the raw event rate using Feature-wise Linear Modulation. This approach significantly improves classification accuracy on standard datasets without relying on rate normalization or invariance.
Deep Auxiliary Learning via Bilevel Feature Optimization for Label-Efficient Medical Image Segmentation
AI summaryThe paper introduces a bilevel auxiliary learning framework that generates auxiliary segmentation tasks automatically from existing labels, using a feature communication module to transfer knowledge. It achieves consistent state‑of‑the‑art label‑efficient performance on three medical segmentation datasets.
SZED at ArGuard Shared Tasks: Multi-Task Label-Attention and Cross-Modal Fusion for Arabic Hateful Meme Detection
AI summaryFusion of MARBERTv2 and CLIP ViT-L/14 embeddings yields 0.7597 macro‑F1 on Arabic hateful meme binary classification, while a multi‑task label‑attention network with cascade calibration achieves 0.3142 macro‑F1 on fine‑grained multi‑label tagging.
REST: Reliability-Guided Spatial Transcriptomics Supervision for Gene Expression Prediction from Histopathological HE-Stained Images
AI summaryREST trains a model to predict gene expression from HE images by using a reliability‑guided spatial transcriptomics framework; it builds a Biological Program Bank, weights spot alignments by reliability, and reconstructs gene sets, achieving the highest macro correlation against state‑of‑the‑art ST‑supervised methods.
Long-wave Infrared Polarimetric Remote Sensing Image Fusion for Vehicle Detection
AI summaryA Mamba‑based fusion network combines LWIR intensity S0 with degree of linear polarization, using physics‑inspired guidance and bidirectional state‑space scanning to integrate complementary modalities. On a new AIR‑LWIR‑Pol dataset, it achieves 55.0% mAP50 and 29.4% mAP50‑95 for vehicle detection.
Channel-Level Mask Guidance for Slimmable Image Compression
Learned image compression (LIC) has achieved promising rate-distortion performance, but most methods lack scalability for diverse platforms with varying computational and storage constraints. Existing slimmable LIC methods achieve flexible complexity via consecutive-channel…
Robustness of Internal Signals for Hallucination Detection in Vision-Language Models
AI summaryThe paper evaluates internal signal‑based hallucination detectors in vision‑language models and finds that a simple token‑position baseline performs comparably, revealing many methods rely on positional shortcuts and are less robust to early hallucinations. It also shows ensembling complementary detectors yields modest gains.
VitronEdit: Universal Visual Editing across Images, Videos, and 3D Scenes
AI summaryVitronEdit trains a single shared sequence model on image, video, and 3D editing, using progressive multimodal training and geometry-aware post‑training. It achieves state‑of‑the‑art 3D editing accuracy while improving image/video edits, thanks to cross‑modal knowledge transfer.
BUMP: Body-driven Unified Multi-object Physics for HOI Synthesis
AI summaryBUMP is a text-driven framework for multi-object human-object interaction synthesis that combines learned human and manipulated object motions with physics simulation for surrounding objects. It achieves strong performance across single and multi-object benchmarks using heterogeneous supervision.
SciFigBench: A Benchmark for Scientific Method Figure Understanding
AI summarySciFigBench provides 4,352 method figures and 40,670 QA pairs to test dense content, visual relations, and flow tracing in VLMs, showing current models lag in relation comprehension; a training‑free SciFigCue module improves results without parameter updates.
Omni-Woodpecker: Hallucination Correction via Evidence-Prior Rebalancing for Omni LLMs
AI summaryOmni-Woodpecker is a training-free framework for omni large language models that corrects cross-modal hallucinations by orthogonally separating and rebalancing grounding evidence and learned priors in hidden states. It improves reasoning and reduces hallucinations across multiple audio-visual benchmarks without modifying model weights.
LVMem: Image-Augmented Memory for Long-Horizon Question Answering
AI summaryLVMem is an image-augmented view-based memory framework for long-horizon robotic question answering that combines optical-flow frame selection with adaptive image retrieval. It outperforms baseline memory models on the OC-NaVQA benchmark without requiring explicit 4D scene graphs.
Counting Contributions, Not Views: Occlusion-Aware Densification for 3D Gaussian Splatting
AI summaryOcclusion‑aware densification filters Gaussian contributions by checking actual visibility, replacing view‑count normalization with average contribution. This reduces noise, improves image quality up to 6% LPIPS, and halves Gaussian count compared to occlusion‑blind adaptive density control.
MatchCityLoc: Aerial Localization in LoD City Models via Surface Normal Matching
AI summaryMatchCityLoc localizes UAVs in LoD city models by extracting surface normals from aerial images with a zero‑shot model and matching them to rendered normals, enabling PnP pose estimation and iterative refinement. It achieves high pose recall and accurate pose with few renderings.
PRIMFLOW: Learning Trajectory Primitives for Few-Step Distillation
AI summaryThis paper introduces PrimFlow, a few-step distillation method that learns continuous trajectory primitives for cumulative displacement in diffusion models. By supervising finite differences with teacher velocities, it improves image generation quality at a single function evaluation compared to existing trajectory baselines.
CARLAR Influence: A Structured Hallucination Risk in Multimodal Cheque Extraction and an Exploratory Compressed-Image Countermeasure
AI summaryThis paper identifies CARLAR influence, a structured hallucination risk where multimodal models erroneously output identical values for distinct cheque amount fields under visual ambiguity, and evaluates a compressed-image few-shot mitigation. It shows that models exhibit high uncalibrated confidence during these errors, while a tabular visual prompting intervention yields model-dependent improvements in extraction accuracy.
Which Depth Estimator for Jetson AGX Orin? A 2026 Benchmark of Monocular and Stereo Methods
AI summaryThis paper benchmarks nine monocular and stereo depth estimators on the Jetson AGX Orin for accuracy, speed, and energy. It finds that stereo methods are significantly more accurate and robust than monocular models, which struggle with scale drift and edge discontinuities.
Language Guidance for Supervised Vision Training: An Empirical Study of Generalization
AI summaryThe paper adds pretrained language model supervision to supervised vision training and evaluates Explicit and Implicit guidance on six generalization tasks. It finds Explicit guidance improves in‑distribution and continual learning, while Implicit guidance helps with shortcuts and adversarial robustness.
On Maximum Clique and Maximum Consensus Formulations in Point Cloud Registration
AI summaryThis paper theoretically analyzes the relationship between maximum clique and maximum consensus formulations in point cloud registration. The authors prove that maximum clique can yield arbitrarily large alignment errors compared to maximum consensus and derive geometric conditions and group-size bounds to tightly control this error.
Generalizing Multimodal Foundational Models in Context-Driven Class Incremental Learning
AI summaryExtends frozen multimodal foundational models to audio‑visual class‑incremental learning by using audio features as contextual guidance for visual representations. Feature‑ and logit‑level distillation reduces forgetting, yielding state‑of‑the‑art performance on audio‑visual CIL benchmarks.
What Not to Erase: Saliency-Constrained Augmentation for Low-Resource Handwritten Character Recognition
Optical Character Recognition (OCR) and handwritten character recognition (HCR) in low-resource historical material collections often rely on data augmentation, yet augmentation strategies are typically evaluated almost exclusively through predictive accuracy. This leaves open…
Steering Vision-Language Models for Visual-Contextual Privacy in Location Disclosure
Vision-language models (VLMs) can infer sensitive information such as geolocation from subtle visual cues, creating privacy risks when responses disclose more detail than the visual context warrants. We study visual-contextual privacy in geolocation disclosure, where a VLM…
Look Before You Reason: Implicit Visual Thinking for Efficient Multimodal Reasoning
Vision-language models (VLMs) have advanced rapidly in multimodal reasoning, yet they remain limited on tasks requiring fine-grained understanding of small or easily overlooked image regions. Recent thinking with images approaches mitigate this limitation by allowing models to…
On the Limits of Scale: Inverse Scaling in Vision-Language Models
Scaling laws suggest that increasing model size, training data, and compute generally yields monotonic performance improvements. While this often holds at the aggregate level, we present systematic evidence that vision-language models (VLMs) exhibit inverse scaling at the…
PerQ: Inverse Generative Modeling for Neural Image Compression via Quantization Error Compensation
Neural image codecs face a fundamental trade-off between pixel fidelity and perceptual quality. Distortion-optimized codecs preserve pixel-wise fidelity but produce blurred reconstructions at low bitrates, whereas generative codecs leverage learned priors to improve perceptual…
Test-time Adaptation for Point Cloud Denoising under Unseen Noise
AI summaryTest-time Adaptive Denoising (TTAD) adapts a pre‑trained point‑cloud denoiser to unseen noise by estimating sample‑specific noise with a residual module, then fine‑tuning via contrastive learning, achieving substantial generalization gains on synthetic and real data.
KPHarness: Co-Evolving Knowledge and Probing for Long-Horizon Gaming Harnesses
A failed gaming trajectory reveals that an agent needs intervention, but not necessarily how its execution harness should change. Effective harness adaptation therefore requires gathering diagnostic evidence, rather than merely proposing alternative recovery rules. We present…
TokenFlow: Latent Token Rectified Flow Matching for 3D Point Cloud Generation
3D point clouds generation requires capturing both part-level layout and surface detail while preserving permutation equivariance. Existing latent generative models often encode an entire shape into a single vector, which limits their ability to retain part-level structure. We…
Using Set Learning to Improve Set-Valued Classifiers on Long-tailed Data
Set-valued classification (SVC) offers a practical approach to communicating uncertainty while maintaining accuracy by predicting variable-sized sets per input rather than fixed-sized sets (e.g. top-1 or top-5). However, set-valued classifiers (SVCs) exhibit significantly…
LoRASpace: A Pool-Wide Shared Substrate for Static and Dynamic Multi-LoRA Composition
Low-rank adaptation (LoRA) attaches small low-rank matrices to a frozen backbone, and public repositories host many LoRAs per backbone, each for a different concept or task. Combining several LoRAs in one generation forces a trade-off. Merging their weights into one update needs…
FocusOVD: Geometric and Contextual Feature Refinement for Open-Vocabulary Small-Object Detection
Open-vocabulary object detection aims to detect both seen and unseen categories by leveraging the semantic knowledge of vision-language models. However, existing methods often struggle with small objects due to insufficient preservation of high-resolution details, ineffective…
Cross-Visibility-Gap Identity Persistence for 4D Scene Graphs
4D scene graphs require persistent object identity when an object disappears inside an opaque carrier and later reappears as a fresh detection. When appearance and target-centric motion are unavailable, identity recovery must rely on relational history. We formulate…
Progressive Subset Selection and Cascaded Consistency for Unsupervised Vision-Language Model Adaptation
Pretrained Vision-Language Models (PVLMs) achieve strong zero-shot recognition but often suffer performance degradation under target-domain shift. Unsupervised adaptation provides a practical way to improve target-domain performance without labeled data, but it remains…
HealBench: A Fine-Grained Benchmark for Evaluating Psychological Comfort in Text-to-Image Generation
AI summaryThis paper introduces HealBench, a 30,000-sample benchmark evaluating whether text-to-image models produce psychologically comforting visuals. Evaluations of 14 models reveal that high aesthetic and semantic quality do not guarantee emotional comfort, and demonstrate that a proposed reward model can improve restorative generation.
Multiscale Temporal-Spectral-Spatial EEG Learning With Modality-Specific and Modality-Invariant Representations for Emotion Recognition
AI summaryGMSI-Net learns multiscale EEG representations and separates modality-specific and invariant features via cross‑modal attention, achieving 96% accuracy on valence/arousal in DEAP and 94% in MAHNOB-HCI, showing EEG plus peripheral signals boosts emotion recognition.
OmniVLoc: Topo-metric Visual Localization for Omnidirectional Cameras
Omnidirectional cameras provide surround-view information, significantly overcoming the field-of-view constraints for traditional visual localization systems. However, existing localization systems fail to fully exploit the characteristics of omnidirectional cameras, struggling…
Less is More: Concise Retrieval and Critic-based Re-ranking for Composed Image Retrieval
AI summaryConcise Retrieval and Critic-based Re-ranking reduces semantic noise by generating minimal, modification-focused captions and re-ranking candidates for consistency and reference preservation, achieving state‑of‑the‑art zero‑shot composed image retrieval performance.
SynCalib: Breaking Dataset Shortcuts in LiDAR-Camera Calibration via Extrinsic-Controllable Synthesis
Online LiDAR-camera calibration remains a fundamental yet challenging task in the field of robotics. Recently, deep learning-based data-driven approaches have emerged, offering high-precision, targetless calibration capabilities. However, existing methods often suffer from a…
CityPaint: A Free Lunch from 2D Diffusion for City-Scale 3D Texturing
Generating high-fidelity, city-scale 3D environments plays a crucial role in applications such as virtual reality, gaming, and urban digital twins. However, the scarcity of real-world city-scale 3D training data forces existing methods to either stitch independently generated…
AnomalyJev: Towards Evaluating Image Anomaly Explanations via Claim-Level Verification
Recent multimodal LLM (MLLM)-based anomaly detection (AD) methods can generate textual explanations that help interpret their predictions. However, reliably and efficiently evaluating these explanations remains challenging. While LLM judges can assess explanations through…
Text-Anchored Modality Experts with Uncertainty-Guided Routing for Multimodal Sentiment Analysis
Multimodal sentiment analysis (MSA) aims to understand user-generated videos by integrating multiple modalities. Existing methods often treat modalities as symmetric, overlooking the fact that language provides a dominant semantic prior while nonverbal signals are complementary…
Real In, Real Out: What If We Only Use Real Data for Scene Text Editing?
Aiming to modify textual content while preserving the original style, Scene Text Editing (STE) has significant practical value in many applications. STE typically requires paired images before and after editing for training. However, recent approaches rely heavily on synthetic…
Causal Spatiotemporal State Evolution for Monocular 4D Reconstruction
While feed-forward models excel in static 3D geometry estimation, extending them to dynamic 4D scenes requires modeling visual geometry as a causally evolving state. Current methods typically rely on unordered temporal aggregation, which lacks the directional structure of…
HiQGS: Hierarchical Spatiotemporal Querying for Feed-Forward Gaussian Reconstruction
Feed-forward Gaussian reconstruction from sparse posed monocular frames requires reasoning over observations that vary jointly in viewpoint and time. Existing methods either predict Gaussian primitives from observation-aligned features or aggregate multi-frame information…
DART: Scaling a Shape Representation Across Anatomical Classes and Imaging Modalities
Current state-of-the-art segmentation models still produce anatomically implausible shapes, such as broken vessels and fragmented glands, on the structures that are hardest to segment. Learned shape representations capture what structures look like and can help correct these…
WaveSpec: Multi-Spectral Iris Presentation Attack Detection via Wavelength-Conditioned Spectral Coherence Adaptation
Multispectral iris presentation attack detection benefits from observing the same iris across multiple near-infrared wavelengths, but existing methods often use wavelength information mainly for conditioning or fusion. This can limit their ability to explicitly model whether the…
Provisioning Robust and Efficient LiDAR Perception for Autonomous Driving
LiDAR-based 3D object detection is a core perception service for autonomous driving, but deployment requires robustness to point-cloud corruption while meeting latency and memory requirements. Existing approaches address robustness and resource efficiency separately, resulting…
PACE-MoE: Adaptive Domain Sampling and Domain-Specific LoRA for Medical Visual Grounding
AI summaryPACE-MoE jointly learns adaptive domain sampling and domain‑specific LoRA adapters for medical visual grounding, routing updates to each imaging domain while preserving the backbone. Across seven domains it boosts mean box IoU by 3.64% over a fixed‑domain baseline.
PLaCA: Learning Patch-Level Future Latent Change for Action Generation
AI summaryPLaCA predicts patch‑level future latent changes from multimodal input and feeds them into an action generator, improving manipulation success: 79.0% on LIBERO‑Long (+5 pp over baseline) and longer sequences on CALVIN (2.78→3.56) by updating <3% of parameters.
Inferring continuous cognitive traits from limited neuroimaging data using pairwise convolutional neural networks
AI summaryPairwise convolutional ranking learns from whole‑brain MRI by comparing subjects, yielding accurate predictions of spatial ability in small samples (N = 108) and outperforming linear and pointwise neural baselines. Backward elimination highlights shared and task‑specific cortical regions linked to performance.
Pairwise Learning Increases Data Efficiency in Deep Neural Networks
AI summaryPairwise learning reframes regression as aggregated ranking comparisons, yielding massive implicit data. It improves data efficiency, enabling a lightweight ResNet10 to achieve 3.64‑year MAE on UTKFace and boosts STL10 classification by up to 5.2% with limited labels.
Reusing Assurance Evidence in AI Pipelines: Validity Conditions, a Fail-Closed Release Gate, and a Replay of 714 Public Data Version Control Pipelines
AI summaryThe paper formalizes when reusing governance‑control evidence in AI pipelines is safe and implements a fail‑closed release gate with Ed25519 signatures. In a replay of 714 DVC pipelines, it finds 3.3% reuse failures and shows static import closures reduce stale reuse to 0.4%.
CovNav: Conformal Prediction-Guided Adaptive Inference for Object Goal Navigation
In object goal navigation (ObjectNav), an embodied agent must find a target object (e.g., "find a bed") in a previously unseen environment, relying on spatial commonsense such as object co-occurrence. Large Language Models (LLMs) encode this commonsense, but deploying them on…
NeoWorld-3: From Pixels to Code-Defined Physical Worlds
AI summaryNeoWorld‑3 reconstructs interactive 3‑D worlds from images using executable, tunable programs, enabling physics‑validated geometry that can be edited by vision‑language agents. It achieves state‑of‑the‑art fidelity and zero‑shot robot execution of planned manipulation trajectories.
A Principled Optimal Transport Framework for Frame Selection in Long Video Understanding
AI summaryOTFS frames long‑video selection as a semi‑unbalanced optimal transport problem, mapping evidence to frames. The method outperforms training‑free baselines on three benchmarks, unifying and generalizing previous query‑aware scoring schemes.
PanoGraphGS: Unifying Panoptic Segmentation and Hierarchical 3D Scene Graphs with Gaussian Splatting
AI summaryPanoGraphGS learns a shared 3D Gaussian representation that yields consistent panoptic segmentation across views and builds a hierarchical scene graph with floor‑room‑object relations. It improves object‑node accuracy by 50% over baselines and sets new 2D panoptic state‑of‑the‑art.
VLA-Retain: Benchmarking Cross-Dimension Effects of Vision Tuning in VLAs
AI summaryThe authors introduce VLA-Retain, a benchmark evaluating how vision-tuning in vision-language-action models affects robustness across diverse perturbation dimensions. They find that targeted fine-tuning frequently causes off-target performance regressions that are not predicted by target gains.
VEOcc++: Learning Persistent Voxel Feature Maps for Embodied Online Semantic Occupancy Mapping and Offline Refinement
AI summaryVEOcc++ learns a persistent voxel feature map that supports online incremental semantic occupancy mapping during exploration and offline refinement afterward. It achieves 67.12 % IoU online and 67.79 % IoU offline on EmbodiedOcc‑ScanNet, setting new state‑of‑the‑art results.
RLSplat: Resource-Linear Budget Control for Feed-forward 3D Gaussian Splatting
AI summaryThe authors propose RLSplat, a budget-controllable feed-forward 3D Gaussian Splatting framework where inference latency and peak memory scale linearly with the Gaussian budget. RLSplat outperforms prior methods across all budgets while achieving significantly faster execution and lower memory usage at low budgets.
When Does Experience Memory Help Closed-Loop LLM Driving Agents? A Controlled Evaluation with Resampled Warmup Pools, Token-Matched Fillers, and Audited Baselines
AI summaryThe study evaluates closed‑loop LLM driving agents with controlled warm‑up experience pools and token‑matched fillers on a HighwayEnv testbed, finding that memory gains depend on the specific warm‑up pool, token fillers dominate, representational depth is ineffective, and baseline quality reverses apparent advantages.
Push-Pull Attentional Anchoring for Diffusion Concept Erasure
AI summaryPush‑Pull Attentional Anchoring removes targeted concepts from diffusion models by shifting cross‑attention representations while bounding cosine similarity, preserving non‑target quality. It outperforms prior methods on Stable Diffusion v1.4 and FLUX across celebrity, style, nudity, and IP categories.
World2HOSI: A Multi-Expert Harness for Long-Horizon Human-Object-Scene Interaction Generation
AI summaryWorld2HOSI orchestrates existing HOI and HSI experts to perform long‑horizon human‑object‑scene interactions using 3DGS scenes and natural language instructions. It achieves higher task completion, interaction quality, and data diversity than single‑expert models.
CILA: Confusion-based Instance-wise Logit Adjustment for Long-Tailed Recognition
AI summaryCILA adapts logits per instance using confusion between target and non‑target classes, restoring informative gradients that class‑wise adjustments suppress; combined with BNI, it improves long‑tailed recognition across benchmarks.
Give the Model a Finger: Developmentally Aligned Evaluation of Counting in Vision-Language Models
Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master counting. Children cross the barrier with epistemic actions: pointing at each item while reciting number words, moving counted…
Revealing Inter-Channel Modality Discriminability: Channel Selective Interaction Fusion Network for Optical-SAR Object Detection
AI summaryThe paper proposes CSIF-Net, a channel-selective interaction fusion network for optical and synthetic aperture radar object detection, which uses frequency decomposition and cross-attention to suppress modality interference and exploit complementary features, achieving state-of-the-art performance on standard datasets.
PixelSR: Pixel-Space Orthogonal Flow Decoupling for Real-World Image Super-Resolution
AI summaryPixelSR introduces a single-stage pixel-space diffusion transformer with orthogonal flow decoupling for real-world image super-resolution. The method successfully eliminates VAE compression bottlenecks to resolve the longstanding tradeoff between structural fidelity and perceptual realism.
How Much Do $\ell_1$ Attacks Need Concentrated Perturbations?
AI summaryThis paper investigates perturbation concentration in L1 adversarial attacks by enforcing per-coordinate caps during optimization. The results show that limiting large coordinate changes retains 97 percent of attack success while significantly reducing peak perturbation magnitudes.
Radar-Based 3D Human Pose Estimation via Motion-Aware Pose Initialization and Observation-Guided Refinement
AI summaryThe paper introduces I2R-Pose, a two-stage framework for radar-based 3D human pose estimation that combines motion-aware initialization with observation-guided refinement. This approach significantly reduces the mean per-joint position error compared to existing methods on benchmark datasets.
Geometry Regularization for Vanilla NeRF via Self-Supervised Cross-Ray Consistency
AI summaryAdds a self‑supervised cross‑ray consistency loss to vanilla NeRF that samples surface candidates, renders virtual rays from different directions, and enforces color consistency, opacity penalization, and normal smoothness. It cuts the mean Chamfer distance by 30.8 % on Replica.
RoboJudge: Multimodal Language Models as Judges for Embodied Video Generation
AI summaryRoboJudge introduces a benchmark and training data for judging embodied video generation, and a fine‑tuned 9B model that achieves a 0.719 Pearson correlation with human ratings—an 0.089 improvement over Gemini‑3.7‑Flash—on Physical Adherence and Instruction Alignment.
Decoupled Representation Learning for Face Presentation Attack Detection under Smart Glass Capture
AI summarySmaGla‑FPAD provides a 297,000‑image smart‑glass face‑attack database with diverse viewpoints and conditions, and DeCoupPAD‑FAS++ uses decoupled representation learning to separate spoof cues from capture variation, achieving the best in‑domain and cross‑scenario attack rejection and bona fide acceptance.
GARRET: Generative Feedback for Region Retrieval in Visual Documents
AI summaryGARRET trains a region-level visual retriever using only (page, question, answer) triples via differentiable generative feedback, avoiding expensive region labels. It achieves 0.4126 macro‑nDCG@10, 2.94 points above baselines, while using 47.8% fewer page vectors.
From Generic to Dedicated: A Novel Optimizer for Online Continual Learning
AI summarySPIN is an optimizer that splits plasticity and stability via averaging, acting as an implicit inverse‑Hessian regularizer for online continual learning. It replaces SGD/AdamW and consistently improves performance across rehearsal‑free and memory‑limited OCL methods.
LaLa-World: Scaling Latent World Models with Natural Language Actions
AI summaryLaLa-World learns a single language‑conditioned latent world model from simulation, robotics, images, and videos, predicting future latent states via a language model. It outperforms domain‑specific LWMs, cuts training cost 6×, speeds planning 13.9×, and enables fast policy inference.
Manifold Prior Guided Deep Unfolding for Hyperspectral Image Reconstruction
AI summaryManifold Prior Guided Deep Unfolding learns a degradation‑free structural prior and injects it into both projection and denoiser modules. The prior‑guided spectral weighting adapts data consistency, while a manifold‑guided Mamba captures global spatial‑spectral dependencies, outperforming state‑of‑the‑art HSI reconstruction methods.
Gaze2HOI: Gaze-Aware Intention Disambiguation for Egocentric Hand-Object Interaction Forecasting
AI summaryThe paper introduces Gaze2HOI, a framework that leverages gaze cues to disambiguate user intent and guide hierarchical egocentric hand-object interaction forecasting. It achieves state-of-the-art performance in both intention inference and motion generation.
S$^3$E-Det: A Strict Spatial-Spectral Equivariant Detector for Multispectral Object Detection
AI summaryS$^3$E-Det introduces strict spatial‑spectral equivariance for multispectral object detection, using SOSC to enforce spectral order, CDEF for pyramid‑aware fusion, and GAH for rotation‑consistent predictions. It achieves state‑of‑the‑art accuracy on MODA, HOD3K, and DroneVehicle with fewer parameters.
SGLDet: Spectral-aware Global-Local Dynamic Modeling for Multispectral Aerial Object Detection
AI summarySGLDet improves multispectral aerial detection by jointly modeling global spectral consistency and local dynamic feature modulation. It uses spectral-driven matching, context memory, and expert‑kernel aggregation to reduce foreground–background imbalance and capture scale/orientation variations, achieving state‑of‑the‑art results on MODA, HOD3K, and DroneVehicle.
EndoWorld-JEPA: Motion-Guided Global-Local Latent World Modeling for Endoscopic Video Representation Learning
AI summaryEndoWorld‑JEPA learns endoscopic video representations by conditioning latent‑state prediction on optical motion: a global predictor models overall state transition while a local predictor weights spatial residual motion. The pretrained model transfers well to four downstream tasks, showing motion‑guided latent learning is effective.
Testing Associative Creativity in Humans and Multimodal LLMs with Difficult-to-Name Visual Stimuli
AI summaryThis paper evaluates associative creativity in humans and multimodal language models using difficult-to-name visual stimuli with varying shape complexity and noise. Results show that humans produce significantly more diverse, novel, and semantically distant descriptions than the tested models.
Staged Continual Adaptation of Multimodal Foundation Models for Japanese Financial Documents
AI summaryThe paper analyzes how staged post‑training (vision‑language alignment, reasoning distillation, domain tuning) affects an 8.4B Japanese financial‑document model. It shows the phases interact non‑additively; the best checkpoint depends on the downstream task.
ICIBS: Increasing the Context for Ice Boundary Segmentation
AI summaryICIBS augments deep learning ice‑boundary segmentation with spatial and geospatial context, improving detection of weak radar reflections in temperate glaciers. It sets a new benchmark on IceAnatomy (10.9 m/43.4 m error) and outperforms baselines by 49.6 % on unseen Patagonia data.
World Models with Configurable Perception
AI summaryThis paper introduces ConfigWM, a world model that uses language task descriptions to filter out visual distractions and extract task-specific state representations. ConfigWM outperforms existing methods by 22.3% on average across diverse noisy environments.
In PRAISE of Attributes: Grounding Retrieval Geometry with Query-Conditioned Subspaces for Person Re-Identification
AI summaryPRAISE adapts vision‑language models at inference by building a query‑conditioned attribute subspace from a bank of prompts, reshaping similarity geometry without labels. It consistently boosts person re‑identification performance on both image‑to‑image and text‑to‑image benchmarks.
Decoupling Action from Egocentric Observation for World Simulation
Egocentric action transfer aims to decouple the semantic action from egocentric observations and reproduce them in new visual contexts, enabling scalable world simulation, embodied policy learning, and robotic data generation. Existing approaches to transferring actions rely on…
Edit in Sync: Unified End-to-End Audio-Visual Editing for Human-Centric Videos
AI summaryUnified end‑to‑end audio‑visual editing (AVE) jointly swaps, inserts, and deletes a person in video and audio without masks, using a synthetic‑to‑real pipeline and a new 100K AVE‑Dataset; it yields better realism, temporal consistency, and lip‑audio sync than separate‑modality baselines.
From Proper Scoring Rules and Edge Biases to Mumford-Shah Models: Robust Unsupervised Hyperspectral Image Segmentation
AI summaryThis paper introduces an unsupervised hyperspectral image segmentation framework based on the Mumford-Shah functional, integrating proper scoring rules, heavy-tailed distributions, and edge biases. The method achieves state-of-the-art accuracy among fully unsupervised approaches on standard benchmark datasets.
Efficient SAM3 Adaptation for Multi-Class Semantic Segmentation via Dense Competitive Representations
AI summaryThis paper introduces DCR-SAM, an efficient adaptation of SAM3 for multi-class semantic segmentation that uses dense competitive representations and decoupled cross-modal interactions. The framework achieves up to a 10.7% performance improvement while delivering a 33 times inference speedup over native SAM3.
Media Intelligence Platform: Multimodal Detection of Synthetic and Manipulated Content
AI summaryThis paper introduces the Media Intelligence Platform, a unified web-based framework integrating specialized detection modules for various types of synthetic and manipulated media. Experiments on benchmark datasets demonstrate reliable multimodal forensics performance across images, video, audio, and text.
Beyond the Frame: Redundancy, Not Trajectory, Carries the Privacy Leak in Video Unlearning
AI summaryThis paper demonstrates that video unlearning methods leak privacy through cross-frame output redundancy rather than object trajectories, allowing simple temporal averaging to recover forgotten classes. To address this, the authors introduce PACE, a new unlearning method that suppresses redundancy using a permutation-invariant adversarial discriminator and temporal consistency objectives to effectively block reconstruction attacks while preserving model utility.
Recast the Past: Stable Long Video Generation with Full Context Memory
AI summaryReCam introduces a memory mechanism for diffusion transformers that compresses historical video latents into vectors while preserving full-context access via native self-attention. It achieves state-of-the-art long video generation quality and superior consistency on VBench-Long.
Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds
AI summaryMatrix‑Game 4.0 decouples a lightweight game engine from an autoregressive video renderer, enabling 480p, 60 FPS real‑time gameplay with 112 ms key‑press latency on a single RTX 5090. It achieves 83 FPS steady‑state throughput and supports natural‑language scene creation and gameplay modification.
Beyond the Evidence Boundary: Behavioral Mirage Under Progressive Degradation in Medical VLMs
AI summaryMedical VLMs often keep giving confident answers even as image quality degrades; the study defines an evidence boundary where models cease to be reliable and shows that MedGemma’s post‑boundary behavior varies from correct recovery to persistent mirage, highlighting the need for boundary‑aware evaluation and better abstention.
Differentiable Task-Driven Coded Aperture Camera Design with 3D Gaussian Scene Representations
AI summaryThis paper introduces an end-to-end differentiable framework for jointly optimizing complex camera optics and downstream perception tasks using a 3D Gaussian Splatting scene representation. The method successfully designs superior coded-aperture depth-from-defocus cameras, outperforming traditional and non-differentiable design baselines.
AdaptMIL: Domain Adaptation for Multiple Instance Learning in Digital Pathology
AI summaryAdaptMIL trains PFMs for digital pathology by first aligning representations across hospitals, then learning MIL aggregation, avoiding interference. It boosts cross‑hospital accuracy, outperforming existing adaptation methods on several whole‑slide image datasets and tasks.
P$^3$-VLM: A Point-based Alternative for Grounded 3D Vision-Language Models
AI summaryThis paper exposes that 3D vision-language models exploit bounding box prompts as semantic shortcuts rather than using visual data. To fix this, the authors introduce a point-based grounding protocol and a novel detector-free model, P3-VLM, which achieves state-of-the-art accuracy while improving spatial reasoning and robustness.
Spherical Geometry-Aware Panoptic Scene Graph Generation for Equirectangular Panoramas
AI summaryThe paper introduces a spherical geometry‑aware framework for panoptic scene graph generation on equirectangular panoramas, fusing perspective predictions and using spherical‑area‑aware relation reasoning. It achieves 36% PQ and 53% mR@50 gains over baselines.
A Process-Informed, Photorealistic Digital Twin for Vision-Based Inline Inspection of C-SMC Strip Deposition
AI summaryThe paper builds a photorealistic digital twin for carbon‑fiber SMC strip deposition, generating 16,000 synthetic images with exact labels. Models trained on this data achieve 0.91 mAP50 on synthetic sets and transfer well to real manufacturing images.
Assessing Representation Semantics Under Language Pretraining in Fine-Grained Classification
AI summaryThis paper introduces an explainability-driven framework to compare vision encoders trained with classification versus caption supervision. The main result shows that caption supervision yields representations with better part localization and lower background reliance, despite similar downstream accuracy.
Auditing Where Concept Bottleneck Accuracy Comes From on CUB
AI summaryThis paper audits Concept Bottleneck Models on the CUB dataset and reveals that their high accuracy often stems from shortcut features and information leakage in internal embeddings rather than genuine visual grounding. The authors show that standard task and concept metrics are insufficient, and CUB's noisy annotations preclude overinterpreting accuracy gaps.
Slot-JEPA: Object-Structured Prediction Interfaces for Latent World Models
AI summarySlot-JEPA decomposes joint-embedding predictive architectures into object-structured semantic slots and spatial masks using optimal transport. This design improves feature transferability across visual benchmarks and enhances goal-conditioned planning in robotic environments.
MoDDiT: Simple Token Routing in Diffusion Transformers
AI summaryMoDDiT learns to route image tokens between MLP and self‑attention, enabling dynamic compute allocation that processes only about 12.5 % of tokens. This reduces training and inference cost while improving generation quality over dense Diffusion Transformers.
WildAIGC-Bench: Can AI Image Detectors Handle the Wild? A Benchmark of Generated and Edited Images in Online Circulation
AI summaryThis paper introduces WildAIGC-Bench, a benchmark of 407 human-verified, real-world AI-generated and edited images gathered from online media. Evaluation of current detectors and multimodal models reveals significant performance drops on these uncurated, authentic online images.
LatticeRank: Spacing-Ranked Consistency for Native-Grid 3D Medical Image Segmentation
AI summaryLatticeRank trains a single 3D medical segmentation model to handle varying voxel spacings by aligning predictions in physical space and weighting consistency loss toward the coarsest spacing. It gains up to 0.034 Dice on coarse grids versus mean consistency while keeping a single‑forward inference.
Reliability-Guided Vision-Language Decoding with Text-Aware Skip Refinement for Medical Image Segmentation
AI summaryRELiF fuses medical text with images using a reliability‑guided decoder and a text‑guided skip refinement, adapting cross‑modal updates based on confidence. It achieves competitive segmentation on radiographs, CTs, and endoscopies while providing uncertainty signals correlated with errors.
PointForward V2: Point-Aligned Interaction for Feedforward Driving Reconstruction
AI summaryPointForward V2 enables scalable 3D query interaction for driving reconstruction by routing information through Static Scene Registers for background and Dynamic Instance Registers for moving objects, using 4D positional encoding. It outperforms prior methods on Waymo, nuScenes, and Argoverse 2.
GAMF-Net: Grouped Availability-Aware Multimodal Fusion for Smartphone Skin Lesion Classification
AI summaryGAMF-Net fuses smartphone skin images with grouped clinical variables, weighting only available groups and normalizing by their availability. On PAD-UFES-20 it improves accuracy to 74.17 %, balanced accuracy 70.11 %, Macro‑F1 68.85 %, and Macro‑AUROC 93.42 %, outperforming all baselines.
OV-Foam: How 3D Representations Shape Semantic Feature Lifting
AI summaryThe paper introduces OV-Foam, which uses a non-overlapping power-diagram scene representation for open-vocabulary 3D feature lifting. This approach simplifies the feature assignment problem and achieves improved segmentation performance over standard Gaussian splatting baselines.
DocLayout-YOLO: Axis-Attentive Document Layout Detector with Mesh-Guided Synthetic Pretraining
AI summaryDocLayout‑YOLO adapts YOLO to document layout detection by training on MeshLayout‑1M, a synthetic corpus of 1M diverse layouts, and adding Page‑Axis Layout Attention to capture horizontal and vertical dependencies efficiently. It achieves state‑of‑the‑art mAP50:95 on four heterogeneous benchmarks, balancing accuracy and speed.
CAPS: Clinically-Verifiable Anatomy-Prioritized Synthesis for Missing MRI Imputation in Brain Tumor Segmentation
Multi-parametric MRI is the gold standard for neuro-oncology, yet acquiring complete 4-modality protocols may be unfeasible in real-world clinical practice. When handling missing modalities, existing automated systems force a critical compromise: robust segmenters operate as…
CrossFit-GS: Out-of-Fold Pose Estimation for COLMAP-Free 3D Gaussian Splatting
AI summaryCrossFit-GS presents a two‑submap framework that cross‑refines camera poses against frozen, unoptimized submaps, rejecting unreliable updates with checkerboard validation. It reduces trajectory drift, improving pose accuracy and novel‑view synthesis compared to prior pose‑free 3D Gaussian splatting methods.
Unified RAW Image Exposure Correction and Restoration via Linear Illumination and Frequency-Adaptive Enhancement
AI summaryRIECR jointly corrects exposure and restores RAW images by first applying a Linear Illumination Correction Network to estimate channel-specific gains, then a Frequency-based Enhancement Network that uses dynamic convolution to denoise or recover clipped textures. It outperforms 16 baselines on a new 223-scene RAW benchmark.
Offline Makes Online Better: Calibrated Cache Compression for Video Generation
AI summaryThe paper proposes CALM, a method that calibrates token merging ratios across layers and heads for autoregressive video generation. By offline profiling the risk map and using dynamic programming online, CALM reduces memory usage while improving video quality, achieving a 21% LPIPS reduction compared to uniform merging.
Attractive Faces Skew Text-to-Image Reward Models and Leaderboards
AI summaryThe paper shows that human annotators rank attractive faces higher than unattractive ones, even when image quality is identical, biasing text-to-image preference models and leaderboard rankings. Attractiveness dominates model scores and can override explicit prompts, suggesting the need to separate quality from cognitive bias.
Learn to Modify: Self-evolving Multi-agent Refinement Framework for Conversational Text-to-Motion Generation
AI summaryThe paper proposes SMART, a multi‑agent system that lets users iteratively refine only the requested spatio‑temporal motion parts in text‑to‑motion generation. It achieves semantic consistency via masked editing and uses a memory‑based strategy to reuse prior edits, improving efficiency.
Region-Guided Food-Package Evidence Extraction and Abstention-Aware Draft R3337 Screening with Selective VLM Review
AI summaryThis paper introduces a region-guided workflow combining OCR, deterministic rules, and selective VLM review to screen food-package images against South Africa's draft R3337 warning criteria. The method achieves 65.0% coverage and 97.4% selective accuracy on a validation set.
Convergent Performance, Divergent Mechanisms in LoRA-Adapted Vision Foundation Models for Brain MRI
AI summaryThis paper investigates how vision foundation models with similar downstream performance achieve brain MRI adaptation through radically different internal mechanisms. The authors show that restricting LoRA adaptation to early-to-mid layers matches full-layer performance while improving parameter efficiency and cross-task robustness.
SplatFlux: Feed-Forward 4D Gaussian Reconstruction and Generation from Video Latents
AI summarySplatFlux decodes video latents into a static‑dynamic 4D Gaussian scene, enabling fast 4D reconstruction and conditional generation. It uses distillation for dynamic masks and motion, achieving state‑of‑the‑art results on 3D/4D reconstruction and generation tasks.
Reliability-Aware Reconstruction of V-Shaped Surfaces with a Low-Cost Multi-Zone SPAD Sensor
AI summaryThe paper introduces a reliability‑aware depth reconstruction method for V‑shaped surfaces using a low‑cost multi‑zone SPAD sensor. By learning point‑wise reliability from sensor confidence, transient peak amplitude, and sharpness, the method reduces mean V‑angle error by 13.8% and point‑to‑plane RMSE by 0.9 mm.
PI-MLLM: Psychophysics-Inspired Multimodal Large Language Model for Oddly Satisfying Video Emotion Analysis
AI summaryPI-MLLM adds psychophysics-inspired alignment of visual motion and acoustic onset dynamics to a multimodal LLM, using MSDA and CAVA to predict viewer emotions in oddly satisfying videos and explain modality contributions. It outperforms prior methods on OSVD’s 12k clips.
VGL-stGen: Learning a shared multimodal model for lung histology and spatial transcriptomics
AI summaryVGL-stGen trains a multimodal model on paired lung H&E images and spatial transcriptomics, learning joint image-expression and expression‑cell‑type mappings. It accurately performs thirteen tasks—classification, gene prediction, and conditional expression completion—by autoregressively predicting gene levels from histology and cell‑type cues.
GALA: Geometry-Aware Language Model for Controllable Object Arrangement
Reasoning about fine-grained placement constraints, such as support and containment, is essential for realistic 3D scene creation. It remains a significant challenge for existing methods that rely on simplified object representations like oriented bounding box (OBB), which are…
ChartEvo: Self-Evolving Agent Harnesses for Long-Tail Chart Reasoning
AI summaryChartEvo is a recursive self-improvement framework that enables multimodal agents to automatically evolve specialized skills and tools for long-tail chart reasoning. It improves the base model's accuracy by over 32% on challenging benchmark datasets.
HRoPE: Homography Relative Positional Encoding
AI summaryHRoPE grounds transformer tokens in local surface geometry by representing each patch as a slanted plane and inserting plane‑induced homographies into attention. The approach learns plane parameters end‑to‑end, yielding state‑of‑the‑art novel view synthesis on RealEstate10K, CO3Dv2, and DL3DV.
CPVB-RG: Cross-Prior Visual Bridging for Prior-Aware Chest X-Ray Report Generation
AI summaryThis paper introduces CPVB-RG, a prior-aware framework for chest X-ray report generation that uses a cross-prior visual bridge to explicitly contrast interval changes against stable anatomy between current and prior examinations. Experiments on MIMIC-CXR and IU X-Ray demonstrate consistent improvements across clinical and language evaluation metrics.
AGGR-Mamba: Mamba-based Network with Attention-Guided Gated Residual Refinement for Undersampled MRI Reconstruction
AI summaryAGGR‑Mamba refines a Mamba backbone’s initial reconstruction of undersampled brain MRI by applying channel‑spatial attention and a gated residual branch to correct aliasing, sharpen edges, and preserve local detail, achieving competitive PSNR/SSIM at R=4 and R=8.
Towards Consistent Geospatial Representations: Aligning One Encoder to Many Foundation Models
AI summaryThe paper introduces Pangaea-Latents to evaluate five geospatial foundation models and proposes GeoMixture, a multi-representation alignment framework. The resulting single encoder matches the best foundation model's average rank, successfully trading peak performance for consistent cross-task generalization.
ABACUS: A Benchmark for Image Editing Models to Assess Object Category and Counting Under Scene Conservation
AI summaryABACUS benchmarks image‑editing models for adding, removing, or changing a specified number of objects while preserving the rest of the scene, using a judge‑free, instance‑level check against panoptic ground truth. It shows that current models fail increasingly with higher counts and complex scenes.
PACE: Plan, Attend, Check, and Emit for Training-Free Response Timing in Streaming Video Understanding
AI summaryPACE is a training‑free framework for streaming video QA that plans observations, checks stability, grounding and completion, and emits answers only when confident, improving PAUC and accuracy by up to 5.9 and 6.2 points over prior methods.
HistoLoop: Iterative Contextual Inference with a Looped Transformer for Histology-Based Spatial Gene Expression Prediction
AI summaryHistoLoop uses a looped Transformer that iteratively updates spot states and refines spatial context at increasing resolutions to predict gene expression from H&E images. It achieves the best macro‑averaged Pearson correlation on HEST‑Bench HVG50 across six cancer tasks.
Auto-Director: Leveraging Multi-Agent Collaboration for Autonomous Video Editing
AI summaryAuto-Director is a multi-agent framework that mirrors professional video editing workflows through collaborative planning, retrieval, assembly, and critique agents. Experiments show it significantly outperforms existing methods in automated and human evaluations across a new benchmark for autonomous video production.
Gated Prototype Transfer for Zero-Support Medical Vision Language Model
AI summaryThis paper introduces GProtoT, a training-free module that transfers visual residuals from supported classes to zero-support classes using clinical ontologies and an evidence gate. It improves missing-class recall and overall accuracy in few-shot medical vision-language adaptation while minimizing negative transfer.
RayCo4D: Ray-Conditioned Continuous Motion Model for Monocular 4D Dynamic Reconstruction
AI summaryRayCo4D introduces a ray‑conditioned continuous motion model that decouples visual data from latent 3D motion, using shared SE(3) motion nodes and white‑noise‑on‑acceleration priors to interpolate motion across fragmented tracks. The method yields more consistent 4D reconstructions, higher‑quality novel views, and continuous‑time rendering.
S-Agent 2.0: Spatial Reasoning with a General Coding Agent
AI summaryS-Agent 2.0 equips a general coding agent with a 3‑D perception toolkit and a persistent scene memory, letting the VLM acquire, store, and interpret spatial evidence in plain language. It boosts Qwen 3.8‑27B by +13.9/+20.5 points on MMSI‑Bench/ReVSI and Gemini‑3.8‑Flash by +11.2/+9.6, achieving state‑of‑the‑art scores.
CountVideoBench: A Benchmark to Evaluate Prediction Biases on the Object Counting task by Video-Language Models
AI summaryCountVideoBench evaluates video‑language models on object counting by presenting 2,600 synthetic videos with balanced counts from 5 to 50; none of the 16 models studied achieves over 27.6% exact match, and accuracy patterns vary, revealing prediction biases rather than pure difficulty.
RACE: Learning Multimodal Search through Referential Abstraction and Compositional Evidence
AI summaryRACE creates multimodal search tasks that force visual reasoning by replacing identity clues with referential or relational references and linking multiple evidence pieces. Fine‑tuning on 12K RACE trajectories boosts Qwen3.5‑27B accuracy on ten benchmarks from 53.5 % to 56.0 %.
VisualEvolve: A Skill Self-Evolution Framework for Visual Reasoning
AI summaryThis paper introduces VisualEvolve, a self-evolution framework that iteratively refines external visual operations and tool-use strategies for vision-language models without updating model parameters. Across diverse benchmarks, VisualEvolve consistently improves model performance while requiring significantly fewer trajectories than standard reinforcement learning methods.
StreamEval: Evaluating Interactive Streaming Video Generation Models
AI summaryStreamEval is a benchmark for interactive streaming video generation with 300 scripted test cases across five tasks. It evaluates task execution, state consistency, visual fidelity, and audiovisual coordination, showing that high visual quality does not guarantee reliable request execution.
A Multi-Agent Framework for Automated Machine Learning with Evidence-based Planning for Medical Image Analysis
AI summaryMedicalMLOmni is a multi‑agent AutoML system that uses evidence‑based planning and parallel hyperparameter search to automatically select medical imaging architectures via a natural language interface. It achieves the best score on 19 of 20 benchmarks, proving the framework’s effectiveness.
Brain2Vec: A Self-Supervised 3D Foundation Model for Structural Brain MRI
AI summaryBrain2Vec pre‑trains a self‑supervised 3D neural network on 74,425 T1‑MRI scans, yielding embeddings that improve brain‑age prediction accuracy and speed convergence versus random init, and outperform volumetric features on most downstream tasks.
Code2Video-Compass: A Comprehensive Benchmark for General-Purpose Code-Based Video Generation across T2V, I2V, and V2V
AI summaryThe paper introduces Code2Video-Compass, a benchmark for evaluating code-based video generation across text-to-video, image-to-video, and video-to-video tasks using a dual-level evaluation framework. Experiments with 15 multimodal models reveal that even leading models struggle to generate complex, high-quality videos reliably.
DISTRACT: Degradation-Indexed Study of Textual Reference Across Competing Targets
AI summaryDISTRACT evaluates how textual descriptions influence vision‑language trackers under visual degradation. The study finds that only Molmo2 and JointNLT use text to guide tracking, with decodable effects at early decision stages, while other trackers show negligible or non‑significant text influence.
WorldSR: Harnessing World Knowledge Search for Grounded Image Super-Resolution
AI summaryWorldSR grounds image super‑resolution on external visual knowledge by decomposing images into entities, retrieving aligned reference images, and using them in a multi‑reference SR model. It achieves higher structural fidelity and visual quality than prior generative‑prior methods.
EvaOmni: An Evidence-Grounded Benchmark for Audio-Visual Reasoning
AI summaryEvaOmni is a 392‑video benchmark that requires models to return both an answer and the exact audio‑visual evidence intervals, using new coverage and bottleneck grounded scores. Gemini3.7‑Flash tops the set with 79.4% accuracy, but still lags in precise evidence localization.
LocaLeJEPA: Maximum Entropy Representations under Geometric Priors
AI summaryLocaLeJEPA extends LeJEPA by replacing the isotropic Gaussian target with a locally Euclidean distribution that respects non‑Euclidean embedding geometries. On ImageNet, it yields higher linear and k‑NN accuracy for hyperbolic ViTs, equivariant ResNets, and standard ViTs.
Answering in the User's Variety: Calibrated Accommodation Pairs for Post-Training
AI summaryThe paper examines how post‑training updates (SFT, DPO, RL) shift chat models from mirroring user dialects toward standard English, finding that DPO still prefers British/Australian spelling over AAE or Singlish. Introducing calibrated accommodation pairs (CAP) restores dialect matching without harming standard‑English users.
CLOSER: Evidence-Aware Active Visual Verification for Embodied Object-Goal Navigation
AI summaryThe paper introduces CLOSER, an active verification framework for embodied object-goal navigation that combines vision-language models with closed-loop evidence acquisition. It improves navigation success rates by actively gathering complementary views of candidate objects to resolve false positives before stopping.
Learning from Automatic 3D Gaussian Splatting Edits for Online Bayesian Scene Change Detection
AI summaryThis paper introduces an online scene change detection method that uses automatically edited 3D Gaussian Splatting models to supervise Bayesian evidence fusion. It achieves state-of-the-art accuracy and real-time performance on benchmark datasets.
GazeLens for Efficient Egocentric Video Question Answering with Simple Gaze-Driven Adaptive Token Reduction
AI summaryGazeLens selects frames using HOG and sharpness, then creates gaze‑focused and complementary token summaries that a lightweight gate blends based on the question, keeping both focus and context in two tokens per frame. It outperforms baselines on StreamGaze with fewer FLOPs.
Deep Learning-Based Road Lane Quality Assessment from Egocentric Dashcam Video
AI summaryThe paper introduces a dashcam dataset and three no‑reference metrics—CSI, LCR, RLCS—to evaluate lane‑marking quality. Baselines achieve R²>0.84 and show strong correlation with human rankings (ρ≈0.73 for LCR).
SenseAlign: Understanding Vision and Touch through Language at Test Time
AI summarySenseAlign improves multimodal object recognition at test time by using class text embeddings to weight each modality and correct shared features, yielding robust performance under lighting, contact, and sensor shifts without retraining encoders. The method outperforms baselines on diverse datasets.
Touch-JEPA: Tactile Representation Learning from Hand-Object Video with Limited Tactile Supervision
AI summaryTouch‑JEPA learns tactile representations from hand‑object videos using a self‑supervised masked‑prediction objective, then optionally fine‑tunes with limited tactile labels. It outperforms supervised baselines on contact detection and magnitude estimation for human and robotic hands.
Latent Bridge: From Video Latents to 3D Geometry via Feature Alignment
AI summaryLatent Bridge connects video generative latents with geometric feature spaces, allowing direct prediction of cameras, depth, and point clouds from video latents without RGB decoding. It trains via feature distillation, achieves competitive reconstruction on Co3Dv2/WildRGB‑D, and reuses a single bridge across video, image, and text generation.
VLM-Guided Contact Retargeting and Generative Partner Modeling for Humanoid-Human Interaction
Humanoid-human interaction requires robots to operate under severe partial observability: during close physical contact, the human partner is occluded, truncated, or outside the robot's field of view, causing pose estimation to fail and often necessitating external tracking or…
Copyright Leakage Detection in AI Generated Videos
AI summaryThe paper introduces a benchmark of 50 copyrighted concepts with video samples, and presents a contrastive retrieval framework that treats copyright leakage detection as video-to-image retrieval. It achieves 99.6% accuracy with only one false negative, demonstrating scalable, explainable detection.
CoGaussian: Open-Vocabulary Gaussian-based Collaborative 3D Scene Understanding
AI summaryCoGaussian encodes a 3D scene as sparse, foundation‑model‑aligned Gaussians to enable open‑vocabulary semantic occupancy prediction across multiple agents. On Semantic‑OPV2V, it surpasses single‑agent baselines, demonstrating improved collaborative scene understanding.
Predictive Latents for Autoregressive Language Models: Possibility, Gauge, and Obstruction
AI summaryThe paper analyzes risks when autoregressive language models iterate latent states without emitting tokens. It proves that a width‑k latent with linear transition can match the Bayes token predictor up to a gauge, but shows sharp obstructions: zero latent loss can still hide token errors and vice versa.
OverSeer: Multi-Agent World Models Beyond What Agents See
AI summaryOverSeer adds fixed, auxiliary camera views to a multi‑agent world model, enabling a world‑centric representation that improves temporal and inter‑agent consistency. Experiments in Minecraft, Doom, and real multi‑camera data show higher consistency and reduced token cost using a delta‑VAE.
Continual Concept Personalization in Diffusion Models via Residual-Space Adaptation
AI summaryContinual concept personalization for diffusion models learns new subjects sequentially via residual‑space adaptation, achieving superior concept fidelity and almost no forgetting. The method identifies key layers, allowing 72% fewer trainable parameters without sacrificing performance.
Can a Display Be Used for Rapid LiDAR-Camera Extrinsic Calibration?
AI summaryScreenCalib performs LiDAR‑camera extrinsic calibration using a programmable display that shows AprilTags, allowing a single static pose or multi‑pose joint refinement. It achieves lower translation error than calibration boards and works on various displays and sensor setups.
SceneRig: An Agentic System for Simulation-Ready 3D Scene Reconstruction from Single Images
Simulation is a powerful tool for robotics, but building diverse, simulation-ready scenes at scale remains difficult. Reconstructing scenes from single images is a promising alternative, but visual similarity alone is insufficient: objects must also be recovered, accurately…
WildMAE: Structure-Aware Masked Autoencoding for Wildlife Image Reconstruction
AI summaryWildMAE extends MAE with a decoder-side, structure-aware refinement that uses zero-gated convolutions and gated cross-attention to recover masked patches. It achieves 27.90 dB PSNR, 0.840 SSIM, and 0.160 LPIPS on Snapshot Serengeti, outperforming vanilla MAE.
Learning the 4D World through Point Trajectory Forecasting
AI summaryThe paper proposes VPM, a multi‑view video model that jointly predicts future frames and persistent 3D point trajectories, improving video realism and motion accuracy. Adding point supervision also enhances action‑conditioned robot control, raising manipulation success.
Neural Implicit Energy Optimization for LiDAR Mapping
AI summaryNeural implicit LiDAR mapping is formulated as optimizing a curvature-based energy over the SDF’s shape operator, yielding smoother surfaces while preserving fidelity on million-point scans, with no extra inference cost.
KoalaGS: Complexity-Aware Continuous Level of Detail for Gaussian Splatting
AI summaryKoalaGS introduces a complexity‑aware continuous level‑of‑detail for 3D Gaussian splatting, using pre‑training image complexity to bias Gaussian placement toward poorly reconstructed regions. It retains full‑capacity quality while significantly improving visual fidelity across lower budgets and enabling smoother streaming transitions.
MindView: Benchmarking Generative Multi-View Spatial Intelligence
AI summaryMindView benchmarks language‑guided generation of spatially consistent views from partial multi‑view observations, exposing large gaps in current models; fine‑tuning BAGEL improves scores by 24.3 points, and the training‑free MindCue selector mitigates performance drops in expanded context.
PsycheMME: A Psychology-Informed Benchmark for Social Understanding and Interaction in Videos
AI summaryPsycheMME is a 1,000‑question video benchmark that separates human understanding from response selection, using psychology theories to evaluate grounding, states, needs, decisions, memory, and risk. It shows that existing models score poorly on verified reasoning, especially for timing and need‑judgment completeness.
UniWALM: Unified World Action Language Model for Embodied Robot Control
Motivated by the human ability to integrate experience across language, perception, and action, we introduce UniWALM, a unified world-action-language model that learns video, action, and language in a single shared representation, where each modality is both a conditioning…
Learning Adaptable Dexterous Skills via Hybrid Dexterous Mimicking
Dexterous multi-finger hands hold promise to unlock diverse skills like functional grasps, in-hand reorientation, and dynamic handover that the human-designed world assumes, yet existing learning recipes force a trade-off. Goal-reaching reinforcement learning generalizes to…
SolarRenderer: Few-Step Generative Reference Rendering
AI summarySolarRenderer is a structure‑conditioned video renderer that uses a frozen MiniMax‑H3 backbone with two lightweight LoRAs, achieving realistic 4‑step generation 12.3× faster than 50‑step methods, and real‑time 768p streaming, while generalizing to 11 unseen structural inputs.
AEON: Unifying Video and 3D World Models
Current world models for explicit world simulation typically follow two distinct pathways: (1) video generative models (e.g., Sora), which simulate realistic appearances but fail to preserve 3D consistency; and (2) 3D generative models (e.g., Marble), which generate…
Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary Rigs
AI summaryThis paper introduces a model that jointly generates synchronized video and 3D motion for arbitrary skeletons using a decoupled architecture. The approach outperforms video-first pipelines in faithfully animating diverse rigs and placing them into consistent 3D scenes.