New Computer Vision (cs.CV) papers, Tuesday 20 October 2026

346 papers announced on arXiv. Get them ranked for you, free →

1. Dependency, Compression, and Synergy: A Unified Information-Theoretic View of Multimodal Learning

   Liangjian Wen, Linjie Li, Jiang Duan et al. · arXiv:2609.14421

   Recent advances in multimodal foundation models have intensified the need to understand how different modalities share, preserve, and complement information. Mutual Information (MI), the Information Bottleneck (IB), and Partial Information Decomposition (PID) provide…

2. I$^2$-Med: Interpretable Medical Inference through Visual-Guided Dynamic Logits Calibration

   Md Sajid, Shreeyut Maheshwari, Akshat Mishra et al. · arXiv:2610.23858

   AI summaryI$^2$-Med adds a visual‑guidance module that re‑scores generated tokens during inference, favoring words that are visually grounded and contextually consistent, without retraining the model. On OmniMedVQA, it yields clinically relevant, visually faithful VQA outputs and cuts hallucinations.

3. SAGE-CP: When Marginal Coverage Hides Route-Conditional Failures in Adaptive VLM Ensembles

   Mushir Akhtar, M. Tanveer · arXiv:2610.23864

   AI summarySAGE-CP trains a fixed routing policy for VLM ensembles and calibrates each exit path separately with Mondrian split conformal prediction, guaranteeing finite‑sample coverage per route. It raises worst‑route coverage from 0.623 to 0.887 while cutting calls, latency, and energy by ~60%.

4. Generative Priors for Total-Body PET Reconstruction

   Movindu Dassanayake, George Webber, Denis Prokopenko et al. · arXiv:2610.23870

   Total-body positron emission tomography (TB-PET) enables fast, low-dose, and dynamic imaging over a long axial field of view, but reconstruction remains challenging due to reduced counts, large image volumes, and increased data dimensionality. Generative models offer powerful…

5. Context-Aware RGB-Thermal-Radar Obstacle Perception for Resource-Constrained ASVs

   Pranav Kedia, Edwin Redhead, Heiko Hamann · arXiv:2610.23873

   Reliable obstacle perception is challenging for small autonomous surface vessels (ASVs): reflections and glare confound visible-light cameras, vessel motion shifts the apparent waterline, and darkness removes the dominant sensing modality altogether. Rather than treating…

6. Latent Multimodal Memory Reasoning

   Cai Ke, Xinghao Chen, Yuanchen Bei et al. · arXiv:2610.23890

   AI summaryThe paper introduces Latent Multimodal Memory Reasoning (LMMR), a neuro-symbolic framework that maps heterogeneous visual and textual memories into a shared sparse concept space to construct a dynamic memory graph. Experiments show that LMMR significantly outperforms existing methods in long-term multimodal reasoning and response grounding.

7. EMMA: Event-Driven Multimodal Memory System with Multi-Agent Retrieval

   Ruochen Zhao, Jiazheng Kang, Gezi Li et al. · arXiv:2610.23892

   AI summaryThe paper introduces EMMA, an event-driven multimodal memory system that segments long videos by semantic events and uses multi-agent retrieval across a hierarchical memory structure. This approach improves long video understanding over standard fixed-length clip compression methods.

8. Path-Conditioned Attention Mobility: From Depth to Execution Hubs in SegFormer

   Aleksandr Druzhinin · arXiv:2610.23893

   AI summaryThis paper demonstrates that SegFormer attention heads can be relocated across depths when their computational context is preserved. Consolidating 120 retained heads into fewer execution hubs reduces latency by up to 5.28% on an NVIDIA L4 without sacrificing segmentation accuracy.

9. GeoForward: Geometry-Guided Cross-View Alignment for Feedforward 3D Gaussian Splatting

   Yiming Shu, Peng Sun, Lang Zhang et al. · arXiv:2610.23895

   AI summaryGeoForward uses calibrated cameras and sparse LiDAR to anchor cross‑view correspondence with explicit geometry, then refines with visual features and predicts 3D Gaussian splatting. On nuScenes it achieves lower depth‑consistency error and Chamfer distance than competing methods while maintaining quality.

10. InJePA: Intention-First Latent Planning with Coupled JEPA World Models for Visual Navigation

    Siming Zhou, Bowen Ma, Ruiran Du et al. · arXiv:2610.23896

    Visual navigation requires goal directed planning informed by predictions of environment dynamics. Representative navigation world model methods propose complete action candidates before predicting consequences for selection or search updates. For each candidate, predictions…

11. RhinoNet: Unsupervised Learning for Phenoscaping and Generation of High Dimensional Point Clouds

    Siddharth Viswanath, Hiren Madhu, Eddie Cavallin et al. · arXiv:2610.23898

    AI summaryRhinoNet learns a geometry‑preserving embedding of high‑dimensional point clouds using simplicial complexes and autoencoders, and generates new clouds via latent flow or geodesic interpolation. It outperforms baselines on single‑cell and spatial datasets, producing organized phenoscapes.

12. EAGLE: Entropy-Aware Geometric Learning Enhancer for Partially Polygon-Supervised Object Localization

    Naveen Stewart · arXiv:2610.23900

    AI summaryEAGLE quantifies per‑class shape complexity with an entropy‑based metric and adds a lightweight post‑training head that refines edges and rescales localization scores, yielding about +0.9 mAP50‑95 on COCO for existing detectors with negligible inference cost.

13. Rethinking weakly supervised localization on chest radiographs: position nulls, protocol dependence, and PACE

    Shixin Qiu · arXiv:2610.23902

    Weakly supervised localization on chest radiographs is reported as a single overlap accuracy, and those numbers are not comparable across papers. We measure what one contains. Against position-null controls that keep a readout's own box and move only its position, class…

14. AI-Experiment: an augmented reality art project co-created with artificial intelligence

    Bernhard Egger, Hans Furer · arXiv:2610.23903

    AI summaryThe paper presents a collaborative process where a latent diffusion model, finetuned on an artist’s photographs, generates synthetic artworks that the artist then paints, and an AR layer animates the finished pieces via generative video models. It releases data, code, and an AR demo.

15. LVMH: A Meta-Optimization Harness for Long-Video Agent Memory

    Liwei Che, Benlin Liu, Dai Li et al. · arXiv:2610.23904

    AI summaryLVMH automates memory-policy design for long-video agents via multi-objective code search. It discovers SEEM, a text-first policy that reduces token use 32× and response time 91×, raising accuracy to 65.7%—a 7.7‑point win over prior state of the art.

16. Instance-Specific Classifier-Free Guidance for Text-to-Image Diffusion/Flow-Based Editing

    Wajeeh Atrash, Tom Tirer · arXiv:2610.23905

    AI summaryThe paper proposes predicting an instance‑specific classifier‑free guidance scale to balance image fidelity and prompt adherence. A lightweight model estimates LPIPS and CLIP‑T scores for each edit, enabling fast selection of the optimal CFG scale and improving editing results across methods.

17. Beyond Generator-Specific Shortcuts: Open-World AI-Generated Video Detection via Adversarial Training

    Qi Zhou, Yuanfan Li, Huiqi Deng et al. · arXiv:2610.23906

    AI summaryFORGE trains AIGV detectors with token‑aware adversarial perturbations that suppress generator‑specific shortcuts while preserving forensic cues, achieving 92.62 F1 on 29 unseen generators and outperforming eight baselines by 5.31 points.

18. EgoVine: Post-Training with Branching Experience Compiled from Real-World Videos

    Qinchuan Cheng, Bowen Lyu, Zhantao Gong et al. · arXiv:2610.23907

    AI summaryEgoVine turns fixed egocentric video demos into interactive experiences by compiling them into symbolic environments, enabling agents to recover from mistakes and compare alternative actions. The method jointly trains action generation and consequence prediction, achieving 81.8% success on main tasks and 66.7% on longer compositions.

19. ContraSafe-VL: A Contrastive Safety Preference Dataset for Vision-Language Models

    Di Wu, Yanyan Zhao, Xin Lu et al. · arXiv:2610.23908

    AI summaryContraSafe-VL pairs the same text query with two contrasting images that change a VLM’s safety label, providing contrastive supervision for visual context in safety judgment. Training on it improves safety recognition and reduces over‑refusal while keeping multimodal performance.

20. When the Editing Intent Is Split: A Cross-Modal Jailbreak Attack for Large Image Editing Models

    Xianmin Ye, Zhiyuan Fan, Tianju Liu et al. · arXiv:2610.23909

    AI summaryLarge image editors can be fooled by split-intent attacks that give a safe text prompt while hiding a harmful visual cue; the proposed SIJA attack succeeds on GPT Image 2 (68%) and Qwen-Image-3.0-Pro (91%) and is mitigated by joint intent reconstruction.

21. Grounding Confusable Movie Moments

    Dong-Hee Kim, Seonwoo Choi, Changbeen Kim et al. · arXiv:2610.23910

    AI summaryCoM²‑Bench presents movie‑grounding pairs that differ only in a single detail, revealing retrieval as the main failure point. A training‑free Agentic Window Search improves retrieval by decomposing queries into attributes, raising mIoU from 18.5 to 26.8.

22. Pseudo-Label Foreground Quality Weighting for Domain-Generalized Fundus Image Segmentation

    Lucas Gabriel Telesco, José Ignacio Orlando · arXiv:2610.23911

    AI summaryAdds mask‑derived pseudo‑label quality weights to Noisy Student training for fundus segmentation. The multitask Dice predictor estimates OD/OC reliability, weighting the Student loss, yielding better OC accuracy and overall domain‑generalization.

23. Junction Geometry Survives an Unidentifiable Blur

    Sohail (Neel), Sarkar, Shakuntala Baichoo · arXiv:2610.23912

    AI summaryThe paper develops a method that recovers the number and direction of intersecting arms in microscopy and X‑ray spectra even when an unknown Gaussian‑mixture blur is present. The estimator attains a minimax rate of \(n^{-1/2}\) and improves arm‑count accuracy and angular error by 24–37 % over existing ridge detectors in three studies.

24. SPIRIT: Sparse Physics-Informed Inverse Rendering for SAR Imaging

    Yimeng Liu, Maolin Gan, Jingkai Lin et al. · arXiv:2610.23914

    AI summarySPIRIT reconstructs 3‑D surfaces from sparse‑aperture SAR by physics‑informed inverse rendering, modeling visibility, phase coherence, and wave propagation. It recovers accurate geometry and signal with over 90% fewer measurements, outperforming conventional super‑resolution baselines.

25. Your Self-Supervised Projection Head Captures Object Co-Occurrence Statistics

    Arthur Aubret, Jochen Triesch, Céline Teulière · arXiv:2610.23915

    AI summarySelf‑supervised models encode object co‑occurrence statistics in their projection heads, while backbones learn categories. By shaping co‑occurrence patterns during pre‑training, category representations emerge faster, showing that leveraging co‑occurrence structure improves SSL learning.

26. Cross-Modal Forgetting and Transfer in Unified Multimodal Post-Training

    Francesco Colasurdo, Ivan Slobozhan, Joaquin Vanschoren et al. · arXiv:2610.23918

    AI summaryCross‑modal fine‑tuning of a unified model (Janus‑Pro‑7B) can cause forgetting in untouched modalities. A single LoRA run preserves other capabilities, but chaining three runs collapses them; anchoring the untrained head prevents this damage.

27. Vocabulary-Incomplete Unsupervised Vision-Language Model Adaptation

    Si Qin, Jiawei Tang, Yaxin Hou et al. · arXiv:2610.23919

    AI summaryVI-UVA adapts vision‑language models to unlabeled data when only a subset of class names is known, using Label Space Reconstruction to build a latent label space and reduce bias toward known classes. LSR achieves state‑of‑the‑art accuracy on eight benchmarks.

28. Flow Matching-Based Semantic Communication With Time-Aware Structural Residual Injection

    Chengyang Liang, Dong Li · arXiv:2610.23922

    Generative semantic communication (SemCom) enables low-bandwidth image transmission by reconstructing images from semantic descriptions. However, text-only schemes often suffer from poor structural fidelity due to the stochastic nature of generative models, while still requiring…

29. SpinDoctor: Finding Innocuous Prompts for Biased Image Generation

    Anna Gerchanovsky, Omer Akgul, Lujo Bauer et al. · arXiv:2610.23925

    AI summarySpinDoctor is a black-box method that discovers seemingly innocuous text prompts capable of reliably inducing specific biases in text-to-image models without degrading image quality. Experiments show it successfully generates biased outputs across multiple models with high consistency.

30. Unbalanced Optimal Transport for Efficient Visual Document Retrieval

    Hoyeon Shin, Jeongyeon Kim, Yeong Jun Koh et al. · arXiv:2610.23929

    AI summaryCompresses VLM visual tokens for document retrieval via an Unbalanced Optimal Transport formulation that balances importance, redundancy, and coverage. The method theoretically bounds retrieval error and outperforms state‑of‑the‑art baselines on ViDoRe V1 and V3.

31. Trimodal Context-Aware Micro-Expression Recognition in Real-World Conversations: A Dataset and a Hierarchical Fusion Network

    Junbo Wang, Yan Zhao, Shigang Wang et al. · arXiv:2610.23930

    AI summaryTrimodal TriCoME dataset pairs micro‑expressions with concurrent audio‑text context and labels each context as external or internal. HFFNet hierarchically fuses fine‑grained acoustic and semantic streams, compresses dialogue, and routes samples between visual‑only and multimodal branches, improving real‑world micro‑expression accuracy.

32. STLQ: Spatio-Topological Linkage Quantization for Approximate Nearest Neighbor Search

    Kuizhi Liu, Junying Zhang, Xindi Ma et al. · arXiv:2610.23931

    AI summarySTLQ is a shallow multi‑codebook quantizer that links vectors within each IVF cluster by a directed‑forest, re‑encoding residuals relative to parent reconstructions. It achieves higher recall than other shallow MCQs at the same bit budget while maintaining linear ADC complexity.

33. Toward AI-Driven Phytoplankton Event Detection from Image Classification: Temporal Aggregation Under Domain Shift

    Ali Rohan, Hassan Shayan, Han-Gue Jo et al. · arXiv:2610.23932

    AI summaryAI image classifiers can reliably generate temporal occurrence rates for phytoplankton taxa. In cross‑domain tests, EfficientNet‑B0’s aggregated predictions matched expert time series (MAE ≈ 0.043, Pearson ≈ 0.97), enabling monitoring and early‑warning potential.

34. VERDICT: Visual Extraction-Result Decoder with Integrated Confidence and Traceable Evidence

    Harikrishnan P M, Goutham Vignesh, Siva Prasad et al. · arXiv:2610.23933

    AI summaryVERDICT attaches a Vision–Language Model to any extractor to localize evidence and calibrate confidence for both extracted values and evidence locations. Fine‑tuning the VLM boosts value AUROC to 0.914 and evidence AUROC to 0.906, enabling high‑precision coverage at 0.80 confidence.

35. CrowdH-Edit: Reference-Guided Instruction-Based Synthetic Hajj Crowd Image Generation for Crowd Counting

    Amirah F. Alshammari, Bander A. Alzahrani, Nahed A. Alowidi · arXiv:2610.23945

    AI summaryCrowdH-Edit generates synthetic Hajj crowd images guided by real references and textual instructions, producing 6,122 clean samples and 387 annotated ones. Adding the synthetic set to training data lowers MAE in three models and RMSE in two, showing reference-guided synthesis benefits crowd counting.

36. DS-Net: A Dynamic Split Network for Edge-Cloud Collaborative Inference

    Xinyun Zhang, Li Wang, Lei Jin et al. · arXiv:2610.23949

    AI summaryDS‑Net enables a single DNN to flexibly switch split points between edge and cloud, cutting memory use by 74.6% and switching latency to milliseconds while a hierarchical compensation mechanism improves accuracy and reduces latency by 42.5% on COCO at 10 Mbps.

37. DISCERN: Identifiable Disease Evidence for Generalizable Medical Multimodal Learning

    Xiaoguang Zhu, Pengyi Jiang, Juncen Guo et al. · arXiv:2610.23952

    AI summaryDISCERN learns a latent disease-state representation that is invariant to cohort selection, removing enrollment shortcuts. It improves state AUROC from 0.043 to 0.883 on synthetic shifts and increases cross‑database mortality AUROC by 0.036–0.035 versus empirical risk minimization.

38. Functionalizing AI-Generated 3D Shapes via Interior Geometry Construction

    Haoyu Tang, Fancheng Gao, Kan Chen et al. · arXiv:2610.23953

    AI summaryWe construct interior geometry for AI‑generated 3D shapes by building an interior‑aware part graph from VLM‑segmented parts, then fitting each part to a generalized cylinder via optimization, producing functionally valid models that match the original exterior.

39. Same Space or Not: Benchmarking Spatial Intelligence Through Cross-View Contradictions

    Huizi Cao, Guanyu Xu, Zhixuan Liang et al. · arXiv:2610.23959

    AI summaryThis paper introduces Same Space or Not (SSoN), a benchmark for evaluating whether vision-language models can maintain coherent 3D scene representations by identifying cross-view contradictions. The results show that current models struggle significantly with this task, achieving low accuracy compared to humans, though performance improves with targeted supervision.

40. A Configurable ETC2 Encoder with Learned Search

    Ziyu Guo, Haomiao Jiang, Sergey Makeev et al. · arXiv:2610.23961

    ETC2 fixes the texture representation and bit rate, leaving the encoder to trade search time for distortion. We present a configurable CPU encoder in which rules or small neural networks allocate conventional ETC2 searches; these searches construct every encoded block.…

41. Generalizable fetal brain biometry via adaptive artificial intelligence

    Yijin Li, Mingxuan Liu, Yi Liao et al. · arXiv:2610.23966

    AI summaryBioTTA adapts pretrained fetal brain biometry models online to unlabeled target MRI volumes, achieving 1.1–2.3 mm MAE and 82–95 % diagnostic accuracy across cross‑sequence, cross‑center, and cross‑nation datasets, matching radiologist performance.

42. Geo-vision: Native-Resolution Geometry and Sampled Topology for Vehicle-Part Damage Classification

    Usef Faghihi, Amir Saki, Serge Jonathan Kogbeti-Wongo · arXiv:2610.23967

    AI summaryGeo-vision classifies vehicle-part damage by fusing appearance, multiscale geometry, orientation, and sampled digital topology in a fixed representation, achieving 0.7577 AP—better than a YOLO11n baseline—while using fewer labels and modest CPU memory.

43. IMPULSE: Push First, Then Predict in V-JEPA 2 Feature Space

    Tingan Jin · arXiv:2610.23968

    AI summaryIMPULSE introduces an active interaction strategy for V-JEPA 2 that uses a few short physical pushes to update latent world models. By incorporating this experiential memory and a Fourier pose head, the method reduces one-second feature-forecast error on novel objects by up to 16.6%.

44. Learning When to Trust and Reobserve Sparse-Frame VideoQA from Temporal Sampling Perturbations

    Shujia Liu · arXiv:2610.23969

    AI summaryPhaseGuard+ probes a frozen VLM at different frame offsets to separately predict answer reliability and the expected accuracy gain of re‑observing or using a stronger model. For Qwen3‑VL‑8B, reliability AUROC improves to 74.8%, and routing 30 % of questions to a larger VLM boosts accuracy by 4.2 pp.

45. Spatio-Temporal Detection of Photovoltaic Systems on Industrial Building Rooftops Using Satellite Imagery and Machine Learning

    Frederick Bruch, Christian Knoth, Florian Detsch et al. · arXiv:2610.23970

    AI summaryThe paper builds a Random Forest model that detects industrial rooftop PV arrays using open-source Sentinel‑2 multispectral data combined with OpenStreetMap building outlines, and implements a web prototype that delivers up‑to‑date detections with quality flags for users.

46. Super-resolution as Spatial Downscaling: Coherence, Structure and Task Utility in Earth Observation

    Bright Wiredu Nuakoh, Ebenezer Afrifa-Yamoah · arXiv:2610.23972

    AI summaryLearned satellite super‑resolution can be viewed as spatial downscaling; across 735 tiles, lightweight models outperform bicubic by ~1.4 dB, recover ~30 % of high‑frequency spectral power, and change land‑use accuracy by up to ±3.1 pp, though they may amplify spurious changes.

47. How Much Can We Prune? Exploring the Performance Limits of Visual Token Pruning

    Janghoon Cho, Shubhankar Borse, Reza Pourreza et al. · arXiv:2610.23974

    AI summaryThe paper formulates visual token pruning as an offline utility maximization problem for image understanding, introducing QA‑Mini‑Batch Greedy (QA‑MBG) to efficiently build nested token subsets. QA‑MBG achieves higher performance than existing baselines, closing the gap to full‑token inference while using far less compute.

48. FrameScout: Scouting Query-Relevant Frames for Long Video Understanding

    Haonan Hu, Shuhao Chen, Weisen Jiang et al. · arXiv:2610.23975

    AI summaryFrameScout selects query-relevant frames for long videos using a streaming selector that incorporates temporal context and a frame-query contrastive objective. It outperforms existing methods on three long-video benchmarks across six VideoLLMs, achieving state‑of‑the‑art results.

49. SPELUNKA: A Framework for Finding Unanticipated Associations in Vision-Language Models

    Cassandra Parent, Isha Agarwal, Kumail Alhamoud et al. · arXiv:2610.23979

    AI summarySPELUNKA automatically discovers unanticipated spurious associations in vision‑language models by generating captions, extracting entities, and using a frequency rule to detect over‑ or under‑representation in retrieval outputs. It achieves 0.93 AUROC across four datasets and outperforms baseline methods.

50. SituHOI: Scalable Capture of Scene-Situated Full-Body Human-Object Interaction

    Qifan Wang, Qihang Fang, Boao Shi et al. · arXiv:2610.23980

    AI summarySituHOI provides 18,429 RGB‑D clips in 16 indoor rooms with whole‑body, object, and scene annotations, enabling joint motion prediction. On a MoCap subset, the pipeline cuts human G‑MPJPE from 195.4 to 68.3 mm and object error from 43.0 to 28.7 mm.

51. Time-JEPA: Grounding Video World Models with Physical Time Series

    Shenyi Li, Ziyang Zhang, Wen Wu et al. · arXiv:2610.23982

    AI summaryTime‑JEPA fine‑tunes a video world model with temporally aligned low‑dimensional physical time series, forcing video representations to track physical state evolution. Across driving, robotics, and human motion, it boosts current‑state decoding, multi‑horizon prediction, and kinematic reconstruction, and improves anomaly sensitivity and generalization.

52. RL Teaches VLMs Where to Attend but Not How to Perceive

    Ziqian Zhang, Junjie Liao, Zhining Zhang et al. · arXiv:2610.23983

    AI summaryReinforcement learning improves VLMs by teaching them to focus on visual tokens rather than enhancing perception. The study shows that RL boosts VQA performance via increased visual attention in LLM layers, recoverable with a simple attention bias plug‑in.

53. Pretrained Features as Teachers, Not States: Compact Latent World Models for Visual Planning

    Yanbo Wang, Nan Lin, Wendelin Böhmer et al. · arXiv:2610.23984

    AI summaryAnchor‑WM uses pretrained DINO features as a teacher to learn a compact latent state that preserves visual detail while avoiding collapse, then discards the teacher for inference. It outperforms other visual world models on four environments and plans 2× faster.

54. BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References

    Xinyu Fang, Yuchi Wang, Weihao Yu et al. · arXiv:2610.23986

    AI summaryBlenderWorld benchmarks autonomous 3D creation from real-world references with 223 tasks across text, image, video, and structured inputs, evaluating models via rendered views and binary rubrics. GPT‑6‑Astra tops the suite at 79.9, but all models fail on fine‑grained spatial coherence and accurate structure.

55. What Generation Knows: Unlocking Fine-Grained Understanding in Unified Multimodal Models

    Tengju Ru, Weitong Lian, Zecong Tang et al. · arXiv:2610.23987

    AI summaryGAFU leverages native generation-side representations in unified multimodal models, compressing them into compact tokens that preserve fine‑grained visual detail. This post‑training framework boosts fine‑grained understanding by up to 20% on VStar‑Bench, outperforming standard SFT baselines.

56. Single-Pass Decision Scores Enable Rapid Classification with Adjustable Thresholds in Medical Multimodal Large Language Models

    Patrick Wienholt, Frederik Hauke, Christiane Kuhl et al. · arXiv:2610.23988

    AI summarySingle‑pass probability renormalization lets multimodal LLMs output a continuous decision score, enabling adjustable clinical thresholds without retraining. On six imaging benchmarks, this method improved AUROC (e.g., CheXpert 0.819 vs 0.777) and achieved 53–79 ms inference.

57. Automatic Cause Determination in Road Scene Understanding Using Qualitative Reasoning and Four-Valued Logic

    Nassim Belmecheri, Arnaud Gotlieb, Nadjib Lazaar et al. · arXiv:2610.23990

    AI summaryThe paper introduces \tool, which uses spatiotemporal reasoning and Belnap’s four‑valued logic on a Qualitative eXplainable Graph to automatically identify complex, possibly contradictory causes for autonomous driving actions, and demonstrates its effectiveness on NuScene data.

58. ProV-Mem: Proactive Memory for Streaming Video Question Answering

    Bingbing Wang, Jing Li, Han Gao et al. · arXiv:2610.23998

    AI summaryProV‑Mem builds a training‑free proactive memory for streaming video QA by decomposing decisions into temporal‑trigger, instructional, and episodic tiers. It outperforms baselines on CAUC and PAUC while keeping low inference latency.

59. Causal Spatio-Temporal Neural Distance Fields for Counterfactual Cardiac Anatomy

    Mathias Lowes, Kurt Butler, Kristine Aavild S{\o}rensen et al. · arXiv:2610.24001

    AI summaryThis paper introduces a causal spatio-temporal neural distance field that integrates structural causal models with implicit neural representations. The method successfully generates realistic 4D counterfactual cardiac anatomy under clinical interventions while preserving subject-specific shape with sub-millimeter precision.

60. spora: A Unified Multimodal Dataset and Platform for Spatial Proteomics

    Benedikt von Querfurth, Johann Wenckstern, Eeshaan Jain et al. · arXiv:2610.24002

    AI summaryspora is a large, harmonized multimodal spatial proteomics dataset of 13,546 multiplex images from 5,477 patients across 35 cohorts and five platforms, with standardized formats, segmentations, metadata, and paired H&E. It also offers a unified interface (spora[io]) and a benchmark suite (spora[bench]) for cell phenotyping, diagnosis, therapy prediction, and virtual staining, enabling fair comparison of foundation models.

61. PACE: Pareto-Adaptive Compression for Efficient Native MLLMs

    Lei Tian, Chunzheng Zhu, Kai Wu et al. · arXiv:2610.24005

    AI summaryPACE compresses visual tokens in native multimodal LLMs by using direction‑aware saliency and Pareto‑optimal budget allocation, achieving near‑lossless performance with only 25–50% token retention and halving FLOPs.

62. Beyond Explicit Intent: Discovering and Shaping Artistic Intent through Human-AI Co-Creation

    Renye Yan, Jialun Zhong, Jikang Cheng et al. · arXiv:2610.24010

    AI summaryThe paper presents a human‑AI co‑creation loop that expands exploration, lets artists discover latent preferences through observation, and then reinforces those preferences via tailored RL, gradually aligning the model to individual aesthetics and revealing vague creative intents.

63. An Anatomy-Aware Staged Pipeline for Vessel-Specific Intracranial Aneurysm Segmentation in CT and MR Angiography

    Gniewosz Drwiega, Wojciech Szymanski, Marek Wodzinski · arXiv:2610.24011

    The TopAneu 2026 challenge requires intracranial aneurysms to be segmented at the voxel level and assigned to one of more than fifty vessel-specific classes, on CT and MR angiography alike. We address it with a six-stage anatomy-aware pipeline rather than a single segmentation…

64. ChairNav: Cross-Embodiment Pre-training and Personalization for Long-Horizon Wheelchair Navigation

    Junhua Huang, Zhizheng Liu, Honglin He et al. · arXiv:2610.24012

    Developing autonomous electric wheelchairs capable of long-horizon sidewalk navigation is critical for accessible personal mobility, yet existing approaches either focus on indoor map-based navigation or rely on expensive, complex sensor stacks to achieve full autonomy. In this…

65. CLARO: Clutter Aware Robot Orchestrator for Open-Vocabulary Grasping using VLM Agents

    Sungsu Kim, Vineet Bhat, Prashanth Krishnamurthy et al. · arXiv:2610.24016

    Robots that follow natural-language instructions unlock a wide range of autonomous applications in households, warehouses, and labs. Real-world environments are diverse and often heavily cluttered, requiring robots to plan over multiple steps, act, and replan as they work toward…

66. DinoMUSE: A DINOv2-guided Framework for Unsupervised Lateral Motion Estimation in Ultrasound Elastography

    Xiyao Meng, Xingyue Wei, Baihua Luo et al. · arXiv:2610.24020

    Quasi-static ultrasound elastography (USE) is a non-invasive imaging method for assessing soft tissue mechanical properties. Limited by low resolution, low sampling frequency and lack of phase information in the lateral direction, accurate lateral motion estimation remains a…

67. ProtoGTX: Prototype-Aligned Graph Transformer for Explainable Whole-Slide Image Analysis

    Iori Azuma, Hiroyuki Kusuhara, Tadahaya Mizuno · arXiv:2610.24023

    Deep learning has achieved strong performance in computational pathology, but explaining whole-slide image (WSI) predictions remains challenging. Existing explanation methods often highlight important regions without identifying their morphological patterns, leaving a gap…

68. SynerGest: Temporal-Spatial Synergistic Gesture Generation for Dyadic Conversations

    Chengxiang Guo, Jian Wang, Lanmiao Liu et al. · arXiv:2610.24031

    AI summarySynerGest generates co‑speech gestures for two‑person conversations, using Contrastive Flow Matching to align semantics and an Impulse‑Trend Synergistic Attention module to separate rhythmic impulses from kinematic trends. The model achieves state‑of‑the‑art realism and fluidity on two public datasets.

69. Implicit Neural Models for Direct Multi-Shell Diffusion MRI Signal Estimation

    Sanna Persson, Fabian Leander Sinzinger, Christoffer Ohlsson et al. · arXiv:2610.24034

    Diffusion magnetic resonance imaging is a noninvasive technique that is sensitive to the motion of water molecules in human tissue. The image signal is commonly described in q-space; however, a complete coverage of this space is often infeasible due to hardware and acquisition…

70. Benchmarking and Improvement of Artifact Segmentation Methods in H&E-Stained Whole Slide Imaging

    Julia Merta, Daniel Mendelsohn, Benedikt Schoser et al. · arXiv:2610.24037

    Technical artifacts introduced during tissue preparation or slide scanning remain a major obstacle to developing robust computational pathology algorithms for hematoxylin and eosin (HE)-stained whole-slide images (WSIs). Despite the growing number of quality-control tools, their…

71. GliomaINK: A Tumor and Anatomy Conditioned Flow Matching Framework for 3D MRI Sequence Synthesis in Glioma Patients

    Moinak Bhattacharya, Annie Singh, Mohammad Chaudhry et al. · arXiv:2610.24044

    Background: Incomplete multiparametric MR examinations limit glioma characterization and downstream AI tools. Generative models can synthesize missing MRI sequences, but anatomical structure and clinically relevant tumor features are challenging to preserve. We propose…

72. LUMINR: Intensity-conditioned Meta-learned INRs for Accelerated Medical Image Registration

    Damian Rynczak, Marek Wodzinski · arXiv:2610.24046

    Implicit neural representations (INRs) have been applied to deformable image registration, where a lightweight network maps a spatial coordinate to a deformation vector. Because an INR encodes one specific signal in its parameters, a fresh network must be optimized for every…

73. Generative Text Modelling and Language-Guided Visual Attention for Deciphering Palimpsests

    Robin Hollifeldt, Tianru Zhang, Jessica Ström et al. · arXiv:2610.24060

    AI summaryAutomated palimpsest under‑text recovery using a custom denoising diffusion model followed by language‑guided handwritten OCR, where CLIP‑derived attention maps provide priors, achieves state‑of‑the‑art reconstruction quality on new synthetic palimpsest datasets.

74. OrTeN: Neural Inverse Graphics for Joint Terrain and Line-scan Camera Trajectory Reconstruction

    Josef Biberstein, Sertac Karaman · arXiv:2610.24075

    AI summaryOrTeN jointly learns a textured height map and corrects orbital line‑scan camera trajectories using neural inverse rendering. On four planetary datasets it achieves terrain RMSEs of 2.35 m (Jezero), 2.82 m (Gale), 16.18 m (Olympus Mons) and 0.43 m (McGetchin).

75. BiNar: Bitstream Native Restorer for Bitstream-corrupted Image Restoration

    Fangcheng Li, Zhen Yu, Kejun Wu et al. · arXiv:2610.24094

    AI summaryThis paper introduces BiNar, a generative framework that restores images from corrupted bitstreams by modeling raw bytes directly without relying on pixel decoding. Experiments show that BiNar outperforms existing image restoration models in both perceptual and objective quality under diverse corruption conditions.

76. Hierarchical and Query-Aware Visual Token Pruning for High-Resolution LVLMs

    Luther Wei, Ling Zou, Kang-Yang Huang et al. · arXiv:2610.24102

    AI summaryPrunes high‑resolution vision tokens in LVLMs using a hierarchical, query‑aware framework that allocates token budgets at instance, crop, and token levels before LLM processing. Achieves 98% and 94.8% performance at 25% and 10% token retention with 2× speedup.

77. Joint, Decoupled, or Agent: When Each Multimodal Alignment Paradigm Wins

    Yi Fang, Gaurav Srivastava, Sirui Ding et al. · arXiv:2610.24108

    AI summaryThe paper compares joint, decoupled, and agent multimodal alignment paradigms across 16 datasets, finding no single winner but that selection depends on supervision budget, task format, and unimodal 5‑NN accuracy, with a diagnostic cascade choosing the best paradigm 71.4% of the time.

78. BLANC: Adaptive Temporal Scheduling for Sign Language Production

    Youngmin Kim, Seoyeong Hwang, Sujung Hong et al. · arXiv:2610.24109

    AI summaryBLANC uses a Multi‑Rate Vector Quantizer to separate body and hand motion into different temporal resolutions and a latent articulatory clock to adaptively schedule variable‑duration motion blocks. It achieves state‑of‑the‑art sign‑language production on PHOENIX‑2014T and CSL‑Daily.

79. HAVANA: A Hierarchical Video-Action Model for Vision-Language Navigation

    Dongzhihan Wang, Yi Du, Chen Wang et al. · arXiv:2610.24122

    AI summaryThe paper introduces HAVANA, a hierarchical video-action model for vision-language navigation that decouples slow language-conditioned future prediction from fast reactive local planning. This asynchronous dual-system design achieves state-of-the-art performance on navigation benchmarks and robust real-world generalization.

80. NV-Generate-Body: 3D Latent Diffusion Model forControllable Anatomy, Tumor, and Demographic Body Mask Synthesis

    Can Zhao, Pengfei Guo, Dong Yang et al. · arXiv:2610.24153

    AI summaryNV‑Generate‑BodyMask is a 3D latent diffusion model that produces whole‑body CT segmentation masks with fine control over organ anatomy, tumor burden, and demographics. Trained on 8,571 masks, it achieves Spearman ρ = 0.70–0.92 for organ size, 85–100 % tumor recall, and strong demographic alignment.

81. SERO: Semantic Prediction for Single-Step Object Removal

    Amreen Hasheem Siraj, Alireza Javanmardi, Pragati Jaiswal et al. · arXiv:2610.24173

    AI summaryThis paper introduces SERO, a semantic prediction framework that achieves single-step object removal and scene completion. By predicting a hidden-region semantic prior in embedding space to guide a latent bridge-matching generator, it produces contextually coherent results in one forward pass.

82. Self-Improving Video-to-Action Models

    Hyunwoo Ryu, Eric Ming Chen, Max Wilcoxson et al. · arXiv:2610.24177

    AI summaryMirror learns video‑to‑action mappings by self‑interaction: it follows a visual plan, records the resulting video‑action pairs, and retrains itself, achieving state‑of‑the‑art performance on Atari, OGBench, DeepMind Control, and Minecraft without prior action data.

83. Precision Cotton Disease Management: Leveraging UAV Imagery and Custom YOLO Deep Learning Framework for Enhanced Detection

    1. Anwar Iqbal*1 aiqbal. rime23smme@student. nust. edu. pk 2. Kunwar Faraz Ahmad Khan1 kunwar. faraz@smme. nust. edu. pk 3. Shahbaz Khan1 shahbaz. khan@smme. nust. edu. pk 4. Khawaja Fahad Iqbal1 fahad. iqbal@smme. nust. edu. pk 5. Sara Ali1 (NUST), 2 sara. baber@smme. nust. edu. pk 6. Saqib Nazir1 saqib. nazir@smme. nust. edu. pk 7. Muhammad Azhar Javed3 muhammadazhar6231@gmail. com 1 Department of Robotics (NUST), Artificial Intelligence (NUST) et al. · arXiv:2610.24182

    AI summaryCotton‑Vision uses UAV images and a lightweight YOLO‑style detector with Ghost Bottlenecks, Efficient Channel Attention, and CSP fusion to detect cotton diseases, achieving 97.7 % precision, 77.7 % recall, and mAP@0.5 of 85.6 %, outperforming current baselines.

84. PREEMPT: Proactive Robustness Evaluation under Editing for Media Protection Techniques

    Qiuyu Tang, Aparna Bharati · arXiv:2610.24205

    AI summaryPREEMPT evaluates proactive image protection methods under sequential editing, normalizing protection effectiveness into survival trajectories and summarizing long‑horizon robustness. It shows that one‑shot tests miss degradation dynamics, revealing distinct failure modes for perturbation and watermark approaches and the pivotal impact of the first semantic edit.

85. Landmark-Free and Scale-Blind Multi-View Registration of Preoperative Liver Models to Intraoperative Surfaces

    Jiaming Feng, Gerardo Loza Galindo, Shahid Farid et al. · arXiv:2610.24218

    AI summarySimCompose jointly estimates full similarity (scale, rotation, translation) from multiple partial liver surface views without landmarks, then refines with a low‑dimensional Neo‑Hookean deformation; it outperforms single‑view baselines, boosting success from 38.4% to 77% on synthetic data and reducing error three‑fold on five in‑vivo patients.

86. Ghost Attack: Using 19th-Century Illusions to Fool 21st-Century Biometrics

    Tyler Fitzsimmons, Adam Czajka · arXiv:2610.24238

    AI summaryGhost Attack shows that a low-cost Pepper’s Ghost projection can simultaneously bypass face matchers, PAI, and morphing defenses, achieving high impersonation success on state‑of‑the‑art systems.

87. Encoder-Free Multimodal LLMs Learn Implicit Encoders

    David Harwath, Desh Raj, Xavier Velez et al. · arXiv:2610.24239

    AI summaryEncoder‑free multimodal LLMs (Gemma‑4 12B, Inkling‑Small) learn “implicit encoders” in their early layers that transform audio and visual inputs into representations largely independent of cross‑modal context. Layer probing shows these implicit encoders match the behavior of standard dedicated encoders, and early‑modality encoding alone matches joint‑encoding accuracy.

88. ULIFLOW: Unified Latent Human-Object Interaction Flow

    Wenqi Cai, Chunzhi Gu, Weichen Zhou et al. · arXiv:2610.24286

    AI summaryULIFLOW models human–object interactions by embedding continuous marker–object relations in a shared latent space, using a Mixture-of-Transformers autoencoder to jointly encode human, object, and relation streams. It achieves longer, coherent interaction synthesis with a blockwise Flow‑Forcing scheme, outperforming prior methods.

89. CRISP: Progressive Illumination Curriculum for Recurrent Low-Light Video Enhancement with Semantic Priors

    Ruirui Lin, Xin Yan Lim, Yini Li et al. · arXiv:2610.24299

    AI summaryCRISP trains a two‑frame recurrent network with a progressive illumination curriculum, moving supervision from interpolated low‑light targets to full references. It achieves top PSNR on DID and SSIM on BVI‑RLV while keeping inference cost low.

90. Hexagonal Basis Projectors for Model-Based Iterative CT Reconstruction

    Ke Chen, Alireza Entezari · arXiv:2610.24363

    AI summaryThe paper introduces hexagonal lattice‐based forward/back‑projection models for fan‑beam CT, deriving analytic ray footprints and detector blur via box spline calculus. Experiments show the linear hexagonal basis outperforms Cartesian and pixel hexagonal models, yielding comparable image and feature accuracy.

91. GazeAgent: VLM-Guided Implicit Gaze Calibration from Egocentric Multimodal Behavior for AI Glasses

    Yuhong Zhang, Khanh An Nguyen, Xinyu Tian et al. · arXiv:2610.24364

    AI glasses typically require explicit, user-specific eye-tracker calibration, and accuracy can drift when the device shifts during use. We present GazeAgent, a VLM-guided framework that continuously and adaptively improves gaze-calibration accuracy from naturally occurring…

92. PACE-Mem: Prefix-Aware Complementary Evidence for Multimodal Agents

    Shidu Ren, Rong-cheng Tu, Hang Zhou et al. · arXiv:2610.24378

    AI summaryPACE-Mem retrieves complementary evidence from multimodal agent memories by using MMR to balance relevance and overlap, then a prefix‑length algorithm selects the optimal retrieval depth. It improves QA accuracy by 1.6% over Adaptive‑k on long‑video benchmarks.

93. Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents

    Wonjun Lee, Vaidehi Patil, Zaid Khan et al. · arXiv:2610.24379

    Recent work shows that attackers can implant dormant malicious behaviors that remain hidden at release but emerge after benign fine-tuning. As foundation models are used as agents, whether dormant poisoning extends to harmful agent actions and multimodal settings remains…

94. Learning from the Right Regions: Class-Conditioned Explanation Guidance for Medical Imaging

    Zubair Faruqui, Rahul Dubey · arXiv:2610.24381

    AI summaryClass-conditioned explanation guidance trains multi‑label medical image classifiers to focus on the correct regions per abnormality, penalizing evidence outside class‑specific annotations. Across EfficientNet, ViT, and Swin models, it boosts All‑Saliency Precision from 0.14–0.10 to 0.40–0.57 with minimal accuracy loss.

95. MSCR: Jointly Balancing Utilization Across Modalities and Discovering Synergistic Information

    Xinyu Chen, Liangjian Wen, Jiang Duan et al. · arXiv:2610.24395

    AI summaryMSCR introduces a regularizer that simultaneously balances modality use and boosts synergistic fusion in multimodal learning. It outperforms MCR on five datasets, improving conditional recovery and cross‑modal attribution while remaining effective with Transformer fusions.

96. Spatons: Spatiotonally Localized Building Blocks for Efficient XR Predistortion

    Mohammadjavad Maheronnaghsh, Sushant Kondguli, Steve Mann et al. · arXiv:2610.24400

    AI summarySpatons are patch‑to‑pixel neural operators that combine tiny MLPs with fixed interpolation for XR distortion predistortion. They achieve comparable SSIM/PSNR to CompenNet++ while using 6× fewer FLOPs, 106× fewer parameters, and 4–10× fewer memory transfers.

97. ZeroDSS: Semantic Segmentation from Class Names Alone

    Tian Ming, Marzi Heidari, Yuhong Guo · arXiv:2610.24405

    AI summaryZeroDSS trains semantic segmentation models using only class names and synthetic data generated via an LLM‑driven prompt distribution and diffusion attention, then iteratively refines masks with feedback loops. It attains 35.0% mIoU on COCO and up to 66.3% on PASCAL VOC.

98. S3D-Mosaic: Subgradient-Guided Cross-Band Structure Transfer for Spectral Demosaicing

    Weiqiang Zhao, Xian Li, Yuzhe Gui et al. · arXiv:2610.24414

    AI summaryS3D‑Mosaic introduces a subgradient‑guided cross‑band structure transfer framework for multispectral demosaicing, using a magnitude‑invariant regularizer and a lightweight DSE‑DConv operator. It achieves state‑of‑the‑art unsupervised reconstruction on CAVE and NTIRE 2022 while generalizing to different MSFA layouts and real‑world UAV data.

99. BodyBridge: Grounded Body Layers for Sim-to-Real Translation of Vulnerable Road Users

    Mark Azer, Fabian B. Flohr · arXiv:2610.24436

    AI summaryBodyBridge introduces Grounded Body Layers, a pixel‑aligned body representation that gives simulated VRUs explicit pose, identity, and mask information for video diffusion models. It improves CARLA‑to‑real VRU pose accuracy by 22 % and reduces body flips by up to 27 %.

100. Do 3D Large Language Models Really Use 3D Evidence?

     Yifei Zhao, Fanyu Zhao, Han Zhang et al. · arXiv:2610.24447

     AI summaryThe paper introduces UNICORNS, a benchmark that tests 3D large language models on grounded reasoning across four evidence conditions, and shows a gap between control accuracy (≈61%) and challenge categories (≈41%). A new post‑training method, HORN, raises challenge accuracy to about 75% without hurting control performance.

101. HumanSkill-4D: Training-Free Language Querying of Animatable Gaussian Avatars in Space and Time

     unsu Kim, Naeun Kim, Dongyoon Han et al. · arXiv:2610.24452

     AI summaryHumanSkill-4D allows users to query animatable Gaussian avatars by language, selecting body parts, garments, or motion intervals without training. It achieves superior 3D/4D referring segmentation over baselines using a multimodal LLM, semantic memory, and motion memory with minimal avatar setup.

102. O-SCD++: Replay-Free Online Scene Change Detection via Recursive Evidence Fusion

     Chamuditha Jayanga Galappaththige, Jason Lai, Lloyd Windrim et al. · arXiv:2610.24462

     AI summaryO‑SCD++ detects scene changes online without replaying past data by recursively fusing evidence onto a fixed 3D Gaussian splat map, achieving 0.575 mIoU and 25.6 fps on PASLCD, surpassing both offline and prior online methods.

103. Unifying Amplitude-Phase Coupling and Reliable Endogenous Priors for Low-Light Image Enhancement

     Tongshun Zhang, Zixuan Zhong, Pingping Liu et al. · arXiv:2610.24467

     AI summaryA joint amplitude‑phase coupling module (APACF) plus a reliability‑prior guided refinement (DGRN) enhance low‑light images by preserving spectral consistency and recovering fine textures, achieving state‑of‑the‑art results on benchmark datasets.

104. RiSE: Residual Subspace Expert for Generalizable Text-Centric Image Forgery Localization

     Kahim Wong, Kemou Li, Yiming Chen et al. · arXiv:2610.24480

     AI summaryRiSE uses frozen SVD components of a pre‑trained ViT and trains independent residual experts on specific forgery types, selecting the best expert at inference with a single forward pass. It beats prior TFL methods by 20%–35% on real data while cutting training time by 10×.

105. Investigating Length Bias and Robustness in LVLMs for Multiple-Choice Question Answering

     Md. Atabuzzaman, Hani Alomari, Chris Thomas · arXiv:2610.24500

     Large Vision-Language Models (LVLMs) demonstrate strong performance on multiple-choice question answering (MCQA), yet their reliability under biased conditions remains underexplored. We investigate length bias, the systematic preference for longer answer options, through…

106. SAP-Depth: Scene-Adaptive Probability Framework for Self-Supervised Monocular Depth Estimation

     Bowen He, Li Zhang, Yujie Chen et al. · arXiv:2610.24506

     AI summarySAP-Depth introduces a scene-adaptive probabilistic framework with a dynamic bins mechanism for self-supervised monocular depth estimation. Experiments on KITTI show that this approach outperforms existing methods in accuracy while better preserving complex edge details and structural recovery.

107. BEM: Learning Bregman Divergences for Concept Erasure in Flow Matching Models

     Mohamed-Hicham Leghettas, Markus Püschel · arXiv:2610.24508

     AI summaryThe paper introduces Bregman erasure matching (BEM), which learns an optimal Bregman divergence via bi-level optimization to remove harmful concepts from flow matching models. BEM outperforms existing baselines in erasing targeted concepts while better preserving unrelated generations.

108. MOMO: Human Motion from Moving Sounds

     Deli Cai, Shuyang Xu, Zhiyang Dou et al. · arXiv:2610.24522

     Human responses to sound depend not only on its semantic content, but also on the evolving spatial relationship between source and listener, conveyed through spatial audio and its temporal dynamics. Modeling human reactions to such dynamic spatial audio is important for…

109. VisConf: Quality-Aware Consensus for Best-of-N in Vision-Language Model Reasoning

     Muhammad Usman, Muhammad Qasim Ayub, Usman Ahad et al. · arXiv:2610.24526

     Best-of-N methods improve text-only LLM reasoning by sampling multiple trajectories before selecting an answer. For Vision-Language Models (VLMs), answer frequency and model confidence alone can be insufficient, as sampled generations also vary in their engagement with the…

110. An Optical Music Recognition Tool for String Quartets

     Jin-Shiang Hu, Bo-Yuan Cheng, Li Su · arXiv:2610.24530

     This paper presents an optical music recognition (OMR) system tailored for PDF scores of string quartet pieces. The approach begins with a SegNet-based segmentation model to localize musical regions, followed by convolutional neural networks and image processing techniques to…

111. Contrastive Hypergraph Source-free Domain Adaptive Object Detection in Adverse Weather Conditions--Extended Version

     Jicheng Yuan, Tung Kieu, Manh Nguyen-Duc et al. · arXiv:2610.24532

     Adverse weather conditions such as fog, rain, snow, and low illumination introduce structured visibility degradations that severely impair object detectors trained on clear-weather data. When source data are inaccessible due to privacy or trans- mission constraints, source-free…

112. Structured Latent Dynamics for Long-Horizon Embodied World Models

     Kevin Chen, Yu-Wei Zhan, Haoran Liang et al. · arXiv:2610.24534

     Reliable long-horizon prediction remains a central challenge for embodied world models. Although recent generative models can synthesize visually plausible action-conditioned futures, their latent dynamics are typically learned through unconstrained transition modules, which may…

113. Write Once, Query Many: Evolving Compact Reusable Memory for Long Videos

     Kerui Chen, Jinglu Wang, Yun Wang et al. · arXiv:2610.24537

     AI summaryVideo‑EMem encodes long videos once into a compact memory of text and sparse images, then retrieves evidence for new queries. It improves Qwen3‑VL‑30B accuracy by ~2.6% on three QA benchmarks while cutting storage 99.86% and per‑query latency.

114. Theory-Guided Instance-Dependent Early Stopping via Second-Order Loss Dynamics

     Ahad Bin Islam Shoeb, Jamal Uddin Tanvin, Kamrul Hasan et al. · arXiv:2610.24540

     AI summaryThis paper introduces Theory-Guided Instance-Dependent Early Stopping (TG-IES), which establishes a theoretical foundation for second-order loss dynamics to identify and skip stable training examples. TG-IES saves over 50 percent of backpropagation instances on CIFAR benchmarks with minimal accuracy loss.

115. Cross-polarization pairing for label-free few-shot thin-section classification

     Bingchen Yan · arXiv:2610.24546

     AI summaryThis paper introduces cross-polarization pairing, a label-free few-shot classification method for rock thin sections that pairs plane-polarized and cross-polarized views of the same microscope field. The approach matches the performance of a supervised model trained on identified rock types without requiring any manual slide labels.

116. Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing

     Yujia Yang, Quan Yuan, Jiawei Fan et al. · arXiv:2610.24590

     AI summaryThis paper introduces Scene-Adaptive VLA, a framework that uses dynamic layer routing based on real-time scene complexity to optimize Vision-Language-Action models for autonomous driving. The approach reduces inference-time computational overhead by 53.5 percent while maintaining driving and language performance.

117. Severity-Channel Conditioning: An Injection-Level Study for Consumer-Device Deblurring

     Abdullah Al Shafi, Kazi Saeed Alam, Sumaiya Rahim Suma et al. · arXiv:2610.24599

     Motion-deblurring models lose substantial accuracy when transferred from professional cameras to smartphone-derived imagery. This paper studies whether a severity descriptor, partly derived from the paired sharp target, helps adaptation on the iPhoneBlur consumer-device…

118. Dynamic self-ensemble on multi-exit residual vision transformer for small-data image classification

     Stephen Ekaputra Limantoro, Yan-Kai Wang, Ching-Chun Huang et al. · arXiv:2610.24602

     AI summaryThis paper introduces ReSEViT, a multi-stage residual vision transformer that uses a dynamic self-ensemble method to improve image classification on small-scale datasets. The approach reduces prediction error and variance while maintaining an efficient model size for data-constrained domains.

119. SCOPE-VSR: A Broad-Coverage Training Dataset for Video Super-Resolution

     Gengyu Liao, Yechun Yong, Zheng Chen et al. · arXiv:2610.24604

     AI summaryThis paper introduces SCOPE-VSR, a diverse dataset of 27k videos curated through semantic rebalancing to address data bottlenecks in open-domain video super-resolution. It also proposes GraCo-VSR, a reliability-aware framework that leverages this dataset to achieve superior reconstruction fidelity and perceptual quality.

120. IntriAdapt: Generalizable Intrinsic-Guided Test-Time Adaptation for Metric Geometry Estimation

     Jiaxing Chen, Zizheng Yan, Jingyu Zhuang et al. · arXiv:2610.24607

     AI summaryIntriAdapt uses camera intrinsics as a geometric verifier to supervise test-time adaptation for metric 3D geometry estimation. By minimizing ray-consistency residuals with a lightweight LoRA adapter, the method improves metric accuracy across diverse benchmarks without needing depth sensors.

121. Beyond Geometric Closure: Measuring and Diagnosing Mechanical Consistency in Visual Dynamics

     Qibin Gu, Yunfeng Yan · arXiv:2610.24619

     AI summaryThis paper introduces a framework to measure mechanical consistency in visual dynamics models by evaluating rigid-chain assemblies for geometric admissibility, actuator effort, and work-energy balance. The results show that current models achieve low visual error while failing fundamental physical and mechanical constraints.

122. Spectral-Aware Proposal Enhancement for Diffusion-Based Multispectral UAV Multi-Object Tracking

     Xu Lin, Tingfa Xu, Tianhao Li et al. · arXiv:2610.24642

     AI summaryThis paper introduces DSATrack, a diffusion-guided framework for multispectral UAV multi-object tracking that integrates spectral features into proposal generation and confidence calibration. It improves the baseline by 4.6 HOTA points on the MMOT benchmark.

123. Modeling Uncertainty Propagation in 3D Morphable Models

     Megh Shukla, Edo Collins, Prashanth Chandran et al. · arXiv:2610.24646

     AI summaryCLUE maps 3DMM parameter variance to per‑vertex uncertainty with a closed‑form formula, reducing computation. It improves mesh and uncertainty accuracy in training and inference, achieving the best uncertainty‑error correlation among baselines.

124. Knowledge Diversion: A Novel Paradigm for Pre-training Decomposable Image Generation Models

     Yucheng Xie, Fu Feng, Ruixiao Shi et al. · arXiv:2610.24664

     Large-scale pre-trained models have become foundational to modern machine learning, yet conventional pre-training objectives primarily optimize source-domain performance without explicitly considering downstream requirements. Consequently, transferable and task-specific…

125. From Views to Worlds: Active Exploration over 3D Worlds for Vision-Language Models

     Qijian Tian, Jiayu Ying, Ke Fan et al. · arXiv:2610.24676

     Vision-language models (VLMs) have achieved remarkable progress in 2D visual understanding tasks, yet the passive view-centric perception paradigm limits their capacity for 3D spatial reasoning. Such reasoning often requires actively acquiring spatial evidence beyond the…

126. MESS: Multi-Exposure Sequence Synthesis for Generalizable Image Enhancement

     Ruodai Cui, Shuaizheng Liu, Rongyuan Wu et al. · arXiv:2610.24680

     Despite the significant progress in image enhancement under adverse illuminations, such as low-light image enhancement, exposure correction, and backlit image enhancement, existing models show limited generalizability to different scenarios. A major bottleneck lies in the lack…

127. Geometry-Aware Interpolation for Inverse Problems

     Pourya Shamsolmoali, Masoumeh Zareapoor · arXiv:2610.24687

     Recovering missing visual information from incomplete or degraded images remains a challenging problem, particularly when spatial variations and complex image structures affect reconstruction accuracy. Although Spatial Transformer Networks (STNs) facilitate spatial feature…

128. Editing Images into Reusable Layers via Compositional Edit-Prior Distillation

     Wen Xue, Yong Liu, Xueheng Li et al. · arXiv:2610.24738

     AI summaryLayerEditor decomposes instructions-guided image edits into reusable RGBA layers using a unified token model and edit-prior distillation. It successfully achieves high editing fidelity while enabling precise layer-level manipulation and multi-layer image decomposition.

129. CorrFlow: Self-Correcting Flow Matching

     Hangwei Zhang, Nathan Rouillé, Chuanxia Zheng · arXiv:2610.24750

     Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them. We introduce CorrFlow, a self-correcting flow-matching framework that reuses this information…

130. CorrCLIPv2: Harnessing Unlabeled Images for Open-Vocabulary Semantic Segmentation

     Dengke Zhang, Fagui Liu, Quan Tang et al. · arXiv:2610.24757

     Open-vocabulary semantic segmentation aims to classify pixels without being restricted to a fixed set of categories. Although Contrastive Language-Image Pre-training (CLIP) demonstrates remarkable potential for open-vocabulary recognition, its segmentation capability is impaired…

131. Design Anything, Design at Scale

     Xuchao Ren, Siyuan Feng, Jing Huang et al. · arXiv:2610.24771

     AI summaryThe paper introduces Design Anything, a graph-based agent harness that preserves design intent and coordinates multi-artifact workflows across diverse domains. It significantly improves constraint satisfaction over baseline models, especially on large-scale design tasks.

132. Evidence Field: Adaptive Covering Programs for Medical Imaging

     Isarar Siddique, Raiyan Siddique, Arshi Niyaz et al. · arXiv:2610.24773

     Medical image classifiers typically return predictions without a native contract for independently reconstructing their spatial decision computation. Evidence Field is a trainable spatial-program architecture that jointly produces class scores, selected covering programs and…

133. ATLAS: Learning to Adapt LVLM Attention for Reasoning Video Segmentation

     Hamidreza Dastmalchi, Ali Cheraghian, Aijun An et al. · arXiv:2610.24783

     AI summaryATLAS learns to transform raw LVLM attention maps into coherent, object‑centric localization cues for reasoning video segmentation, fusing high‑res frame attentions with low‑res video attentions via a lightweight U‑Net with temporal attention. It outperforms frozen‑backbone methods on benchmark datasets.

134. Video-Language Entailment for Video Similarity Estimation

     Benedetta Liberatori, Lorenzo Vaquero, Elisa Ricci et al. · arXiv:2610.24786

     AI summaryVICE estimates concept‑based video similarity by cross‑referencing each video with the other’s caption using a large multimodal model, then fusing the two entailment scores to output similarity and a textual explanation. It outperforms simultaneous‑video baselines on ConViS‑Bench and StepDiff.

135. UniAVC: Unified Audio-Video Creation via Progressive Training and Asymmetric Scheduling

     Kien T. Pham, Yuhao Liu, Chong Hou Choi et al. · arXiv:2610.24791

     AI summaryUniAVC trains a single model with separate audio and video branches, then fine‑tunes on paired data using asymmetric noise schedules and cross‑modal supervision. It achieves text‑conditioned joint synthesis, bidirectional cross‑modal generation, and editing, outperforming prior methods while requiring no task‑specific adapters.

136. Directionally Factorized Illumination with Spherical Harmonic Rotation

     Yimeng Dou, Adam M. Saunders, Andre T. S. Hucke et al. · arXiv:2610.24811

     Many real-world capture scenarios involve an active light source rigidly attached to the camera: camera flashes for dark-environment imaging, underwater vehicle lights, and endoscopic/laparoscopic medical procedures. As the camera moves, the co-mounted light moves with it,…

137. SwitchSplat: Learning Discrete Object States from Casually Captured Scenes

     Yizheng Xie, Yalda Foroutan, George Shramko et al. · arXiv:2610.24812

     AI summarySwitchSplat learns discrete on/off states for interactable objects in casual video captures by decomposing a 3D Gaussian Splatting scene into a shared static base and independent slot Gaussians. It outperforms static/dynamic baselines in novel‑view synthesis and enables rendering unseen state combinations.

138. SL-Ground: Sentence-Level Grounding for Fine-Grained Open-Vocabulary Object Detection

     Yang-Syu Li, Cheng-Huan Yang, Chia-Yung Jui et al. · arXiv:2610.24814

     AI summarySL‑Ground replaces token‑level scoring with a sentence‑level representation for open‑vocabulary detection, directly matching object queries to full category descriptions. It boosts fine‑grained AP from 14.5 to 86.0 on Stanford Cars and outperforms baselines on FG‑OVD.

139. VTF900: A 900-Subject Visible-Thermal Face Recognition Dataset and Benchmark

     Jacob Abaare, Bianca Ianetta, Cedric Nimpa Fondje et al. · arXiv:2610.24832

     AI summaryVTF900 is a 900‑subject visible‑thermal face dataset with expression, pose, and occlusion variations, enabling large‑gallery verification and identification. Benchmarking shows transformer models achieve high AUC (~97.7%) in frontal cases, while domain‑adapted methods (DPIT‑ResNet50) better handle off‑pose conditions, revealing scale‑dependent performance drops.

140. PCHMR: Empowering and Benchmarking Human Mesh Recovery in Privacy-Constrained Real-World Settings

     Zeyu Cao, Qingxuan Wu, Zhiyang Dou et al. · arXiv:2610.24841

     AI summaryThis paper benchmarks human mesh recovery under privacy constraints, training state-of-the-art models with federated and secure‑aggregation methods, and introduces DePoser, a depth‑based local fine‑tuning pipeline that boosts personalization accuracy on in‑the‑wild data without centralizing images.

141. Arbitrary-Scale Anisotropic 3D MRI Super-Resolution with Gaussian Prototypes

     Aditya Rastogi, Ramon Leiser, Prateek Singh et al. · arXiv:2610.24858

     AI summaryGSMR3D is a single-model 3D MRI super‑resolution method that reconstructs any isotropic or anisotropic resolution using learned Gaussian prototypes, scale‑conditioned encoding, and exact measurement consistency. It outperforms geometry‑specific specialists on isotropic super‑resolution and preserves lesion segmentation accuracy across all geometries while requiring only ~6 s per scan.

142. SpriteStage: Long Horizon Animation Generation via Progressive Scene Programming and Replay

     Zhaoqi Wang, Zijian Zhang, Jiamou Liu · arXiv:2610.24869

     AI summarySpriteStage builds long animations by progressively enacting each scene from a checkpointed world state, revising based on runtime feedback, and re‑orchestrating to engine assets. It produces 9.9–14.3 min narratives, showing that iterative scene programming yields coherent, extended stories.

143. Goal Images Are Control Targets: Spatially-Grounded Goal-Image Synthesis for Robotic Manipulation

     Riddhi Chatterjee, Ritish Shrirao, Kushagra Mehrotra et al. · arXiv:2610.24885

     AI summaryThe paper evaluates goal-image generators for robotic manipulation, showing that downstream policy success depends entirely on manipulation adherence rather than general photorealism. It introduces SGE-Goal, a spatially-grounded editing framework that outperforms large proprietary models in multi-step robotic tasks.

144. Multi-Speed Field Evaluation of a Semi-Autonomous Precision Sprayer for Weed Control in Vegetable Crops

     Md Khairul Islam, Boyang Deng, Yuzhen Lu et al. · arXiv:2610.24893

     AI summaryEvaluated a semi‑autonomous tractor sprayer with YOLOv11 and ByteTrack at speeds 0.13–0.45 m/s; at 0.13 m/s, detection 94.6 % mAP and 93.2 % recall, but weed hit rate dropped from 78.3 % to 35.0 % as speed increased, failing commercial thresholds.

145. Predicting Individual Differences in Lipreading From Eye-Tracking and Pupillometry During Silent Visual Speech

     Puneet Kumar, Satyarth Malik, Gao Zheng et al. · arXiv:2610.24898

     Lipreading ability varies widely, including among people who are hard of hearing. Although mouth-directed attention is linked to speechreading outcomes, it remains unclear how much lipreading skill can be inferred from viewing behavior and which signals carry this information.…

146. Stable Signer V3: 4D Deformable Primitive Splatting Is All You Need for Sign Language Production

     Sen Fang, Liwei Che, Dimitris N. Metaxas · arXiv:2610.24926

     AI summaryStable Signer V3 uses 4D Deformable Primitive Splatting with face, body, and tunneling super‑resolution modules optimized once per signer, reducing facial error 71.4 %, body error 35.1 %, and RGB error 24.2 % while enabling interactive high‑resolution sign language avatars.

147. LowNet: Looped Wavelet Neural Networks with 24 Learnable Parameters for Solving Poisson Equations

     Yuanhao Gong · arXiv:2610.24941

     AI summaryLowNet uses a 24‑parameter looped wavelet neural network that solves Poisson equations with high accuracy (up to 100.74 dB PSNR after three loops) and linear‑time complexity, making it compact and efficient for data‑specific applications.

148. Meta Inverse Prompting for Video Generative Models

     Yerin Jung, Jongheon Jeong · arXiv:2610.24965

     AI summaryMeta Video Inverse Prompting (MVIP) learns a meta‑prompt that captures temporal dynamics, enabling fast inversion of videos to textual prompts without iterative optimization. It achieves higher reconstruction fidelity and significant speed gains over image‑based and captioning baselines.

149. Visual Sufficiency Landscapes: Understanding Query-Conditioned Reliability in Visual Token Pruning

     Zhenghan Li, Kexin Sun, Wei Zheng et al. · arXiv:2610.24968

     AI summaryDefines “visual sufficiency landscapes” to measure how many token subsets preserve a model’s output for a given query. Shows significant variation across questions, improving failure prediction and reducing loss by up to 14% on LLaVA and similar models.

150. MultiAltitudeSR: Multi-Altitude Agricultural UAV Image Super-Resolution via Physical Ground Sampling Distance Conditioning

     Xintong Jiang, Yu Tian, Yixue Liu et al. · arXiv:2610.24992

     AI summaryMultiAltitudeSR conditions on source and target ground sampling distances, training on real UAV images from 10–70 m to handle 21 altitude pairs (scale 1.18×–6.03×) without fine‑tuning. It outperforms synthetic‑degradation baselines, achieving competitive LPIPS (0.2236), DISTS (0.2431), and NIQE (8.9020) on the 20‑to‑10 m pair.

151. UtoMe: Observation-Uncertainty-Guided Token Merging for Weather Foundation Models

     Yunbei Zhang, Janet Wang, Xi Xiao et al. · arXiv:2610.25001

     AI summaryUtoMe introduces an observation-uncertainty-guided token merging framework for weather transformers that combines feature similarity with ensemble data assimilation spread. At a 50 percent token reduction, it significantly lowers forecasting error compared to standard token merging while maintaining reduced computational costs.

152. ReTrace: Visual Backtracking for Efficient Multimodal Large Language Models

     Guangzong Si, Dong Wang, Yifan Yu et al. · arXiv:2610.25005

     AI summaryReTrace introduces a dormant token state for MLLMs that decouples early computational deactivation from permanent deletion. This enables the model to retrieve information from previously pruned visual tokens during middle-layer reasoning, achieving a 2.83x prefilling speedup with minimal performance loss.

153. FineSplat: Fine Patches, Local Groups, Detailed Splats

     Hyebin Kim, Hwasik Jeong, Joonhyuk Park et al. · arXiv:2610.25011

     AI summaryFineSplat recovers fine geometry and appearance in feed‑forward 3D Gaussian splatting by using fine image patches, dual locality grouping (view‑wise and 3D spatial), reconstruction‑error feedback, and a compact Gaussian generation scheme, achieving state‑of‑the‑art results with equal or fewer primitives.

154. IECLA: Interpretable Emotion Contribution Learning for Affective Image Editing

     Yijie Guo, Dexiang Hong, Weidong Chen et al. · arXiv:2610.25015

     AI summaryIECLA separates source‑emotion attribution from target‑aware editing, learning signed object‑level contributions via counterfactual interventions and a context‑conditioned predictor. It consistently improves target‑emotion expression while preserving content across six backbones and is validated by VLMs, humans, and user studies.

155. PRET is a few-shot system for pan-cancer recognition without example training

     Yi Li, Ziyu Ning, Tianqi Xiang et al. · arXiv:2610.25054

     AI summaryPRET is a training-free, few-shot system for pan-cancer pathology image recognition. Evaluated on 23 benchmarks, it outperforms existing methods on 20 tasks and surpasses 11 pathologists in lymph node metastasis detection using only eight slide examples.

156. Localization and Neural Splatting on the Edge: An FPGA Scene Memory for Robots

     Que Haohua, Mingkai Liu, Handong Yao · arXiv:2610.25068

     Resource-constrained robots and autonomous vehicles need scene memories that localize a camera and render nearby views without a desktop GPU. Scene-coordinate regression (SCR) and 3D Gaussian splatting (3DGS) fit because they avoid an SfM stage, but ACEsplat-style pipelines are…

157. MoWE: Spatially Adaptive Mixture-of-Wavelet-Experts Pooling for Ultrasound Lesion Segmentation

     Reaz Hassan Joarder, Sidratul Tanzila Tasmi, Alif Arshad Bakshi et al. · arXiv:2610.25074

     AI summaryMoWE replaces max‑pooling in U‑Net with a spatially adaptive mixture of wavelet experts that route each region to a suitable wavelet family based on texture. It boosts ultrasound lesion Dice by ~2 % and cuts parameters by 6.4 %.

158. MoCKEFA: Model-Conditioned Keyframe Extraction and Fair Attribution for Video Model Explanation

     Lahiru Wijayasingha · arXiv:2610.25075

     AI summaryMoCKEFA extracts a few keyframes that preserve a video model’s predictions, then assigns their importance using Shapley values, outperforming per‑frame attribution. Experiments confirm keyframe attribution improves explainability and reveals significant temporal changes between adjacent keyframes.

159. Sufficient Frame Subsets for Video Understanding

     Lahiru Wijayasingha · arXiv:2610.25084

     AI summaryThe paper introduces Sufficient Frame Subsets, compact groups of frames whose masking of all others preserves a model’s prediction, and shows that models with more such subsets perform better. Experiments on UCF101 and Something‑Something v2 confirm the method’s compactness and diagnostic value.

160. Think Visually Across Space and Time: Grounded Reasoning for Video Segmentation

     Junwei Wen, Deshui Miao, Guangming Lu et al. · arXiv:2610.25090

     AI summaryTemporal Chain of Video Primitives (TCoVP) grounds reasoning in video by linking entity references, timestamps, and spatial evidence into a chronological chain. Experiments on five VRS benchmarks show TCoVP outperforms baselines on mask trajectory prediction.

161. LC-TSRNet: A Land-Cover-Guided Two-Stage Network for Annual MODIS Surface Reflectance Time Series Reconstruction

     Wenjuan Zhang, Xuesong Ma, Zhen Li et al. · arXiv:2610.25120

     AI summaryLC‑TSRNet reconstructs missing MODIS surface reflectance by using land‑cover priors in a first stage and multi‑scale spatiotemporal learning in a second stage, achieving 0.56 dB PSNR gain and 0.13° SAM reduction over baselines.

162. HullProbe: Measuring Interventional Part Reliance in Maritime Vessel Classifiers

     Faisal Mohammad, Mohamed Younis · arXiv:2610.25137

     AI summaryHullProbe directly measures how much a maritime vessel classifier relies on each vessel part by swapping hulls or superstructures, using a Walsh transform to compute interventional influence. It shows that even after de‑biasing, about 46 % of the classifier’s decision is driven by the generic hull.

163. Reflection-Guided Dynamics Rectification for Text-to-Image Generation

     Shuo Xu, Hanrui Ying, Debin Liu et al. · arXiv:2610.25139

     AI summaryReflection‑Guided Dynamics Rectification (RGDR) uses reflections of a generated image to intervene at every sampling step, correcting the trajectory of a diffusion model. On fixed budgets, RGDR boosts final checklist scores by 1.51× over TIR and 1.81× over Dynamic CFG.

164. Image Dehazing as In-Context Tabular Inference

     Le-Anh Tran · arXiv:2610.25142

     AI summaryImage dehazing is recast as an in‑context tabular regression: patches of a hazy image are encoded with handcrafted haze descriptors, pretrained visual embeddings, or hybrids and fed to a pretrained tabular foundation model. The model predicts transmission values without fine‑tuning, and the resulting maps yield competitive haze‑free reconstructions.

165. PLANE: Plug-and-Play Fine-Grained Visual Experts for Hallucination Mitigation in Large Vision-Language Models

     Jianghao Yin, Lang Yu, Qianxi Yu et al. · arXiv:2610.25146

     AI summaryPLANE introduces fine‑grained LoRA visual experts and an instruction‑aware gate to enrich CLIP features, reducing hallucinations in large vision‑language models. On new GPT4‑free GIFT benchmark and existing tests, it cuts hallucinations with negligible latency, outperforming state‑of‑the‑art methods.

166. Map2Vid: Embodied Navigation Video Generation with Human-Friendly Motion Control

     Yulin Dai, Yuqian Fu, Tianwen Qian et al. · arXiv:2610.25223

     Controllable video generation has recently achieved impressive visual quality, yet existing control paradigms remain poorly suited to embodied navigation. Text prompts are spatially ambiguous, while camera-trajectory control relies on low-level motion specifications that are…

167. Evaluating Cloud and Self-Hosted OCR Systems for Arabic, English, and French Documents: Bounding Box Detection, Text Recognition, and Mathematical Expression Recognition

     Jumana Alsubhi, Ahmed Alhusayni, Abdulrahman Gharawi et al. · arXiv:2610.25235

     Choosing an OCR system for a multilingual archive requires balancing accuracy against deployment constraints. This study compares cloud and self-hosted systems on three tasks: bounding box detection on 48 Arabic, English, and French pages, text recognition on 315 pages across…

168. GeoV2V: Geometry-Grounded Video Diffusion Model for Driving Scene Generation

     Yuchen Xi, Yuhan Guo, Chenwei Hou et al. · arXiv:2610.25245

     Novel view synthesis under trajectory changes is essential for autonomous driving simulation, yet existing methods struggle to generate consistent videos when extrapolating to unseen viewpoints such as lane shifts. Reconstruction-based approaches maintain geometric consistency…

169. Visual Scaffolding: Toward Native Multimodal Reasoning

     Run Shao, Zhaoyang Zhang, Ziyu Li et al. · arXiv:2610.25247

     The success of chain-of-thought (CoT) in complex reasoning can be attributed to a textual scaffolding effect: explicit intermediate results can be examined, reused, and composed to support subsequent reasoning, much as scaffolding supports a building under construction. However,…

170. Persistent Representational Misalignment Induces Vision-Language Model Failures

     Yonatan Gideoni, Yoav Gelberg, Tim G. J. Rudner et al. · arXiv:2610.25252

     Vision-language models (VLMs) fail at some tasks that are simple for humans---but why? We hypothesize that many such failures stem from VLMs' difficulties translating between vision and language representations, a misalignment the Platonic Representation Hypothesis predicts…

171. PanoDent-KD: From Generalist Knowledge Distillation to Cross-Dataset Specialist Detection and Embedded Deployment in Panoramic Radiographs

     Muhammad Waqas Ashraf, Imad Ali Shah, Ali Hassan · arXiv:2610.25256

     Reliable panoramic-radiograph detection requires consistency across datasets with different class definitions and deployment software. We present PanoDent-KD, an auditable 1280-pixel YOLO26m pipeline linking a 12-class source generalist, feature-level knowledge distillation…

172. GeoART: Geometry-Aware Refinement via Test-Time Scaling for Camera-Conditioned Video Diffusion

     Liudi Yang, Ziyuan Liu · arXiv:2610.25273

     Camera-conditioned video diffusion models often suffer from inaccurate camera control and geometric inconsisten- cies because camera constraints and estimated geometry are only weakly enforced during denoising. Existing reme- dies often require costly geometry-aware fine-tuning…

173. Holo4D: Holistic 4D Reconstruction as Geometric Control for Video Diffusion

     Yushi Lan, Zeren Jiang, Koichi Namekata et al. · arXiv:2610.25276

     Synthesizing a monocular video from a new camera trajectory requires preserving the observed scene geometry and motion while completing regions that become newly visible. Feed-forward 4D reconstructors recover the structure of the observed scene, whereas video diffusion models…

174. Learning Lane-Structured Conditioning Fields for Online HD Mapping

     Andrea Bertogalli, Anthony Dell'Eva, Marcello Ceresini et al. · arXiv:2610.25284

     AI summaryThe paper introduces the Lane Field, a dense BEV-aligned representation encoding local lane structure, and a plug‑and‑play framework that modulates BEV features and guides deformable attention with this field. It consistently outperforms baselines and sets a new state of the art on OpenLane‑V2 lane perception.

175. Vision as Proprioception: Estimating Humanoid Robot Pose from Video

     Jiaxi Jiang, Matthias Heyrman, René Zurbrügg et al. · arXiv:2610.25288

     Estimating a humanoid robot's internal state from video has a wide range of real-world applications but remains unexplored. During interactions with a robot, its onboard state information may be unavailable due to privacy, ownership, or communication constraints. Meanwhile,…

176. The Explainability Illusion: Must ViTs Exhibit Patch-Decoding Interpretability?

     Ronen Schaffer, Hila Manor, Tomer Michaeli · arXiv:2610.25289

     Transformers are widely assumed to exhibit token-decoding explainability. In particular, in vision transformer (ViT) classifiers, decoding individual patch tokens has become a common diagnostic tool due to its tendency to yield human-interpretable outputs. In light of the…

177. MotionShift: Environment-Compatible Motion Transfer via Planner-Evaluator Synergistic Adaptation

     Yanchen Liu, Junyao Gao, Yanan Sun et al. · arXiv:2610.25290

     Existing motion transfer methods excel at exact, rigid motion replication from reference videos to target scenes, whereas they often fail to account for the compatibility between the transferred motion and the target scene environment in accordance with physical laws. In this…

178. SENRI: Understanding the Space-Event Nexus in Real-Life Videos

     Kehan Lan, Kaining Ying, Henghui Ding et al. · arXiv:2610.25309

     AI summarySENRI studies how events and spatial contexts are linked in long egocentric videos, providing a benchmark of 500 multiple-choice questions over 10 hours of footage. Existing memory systems underperform, but a TGVC-based controller, SENRI-Agent, boosts accuracy without changing stored memories.

179. MPIG: Multi-angle Polarization Image Generation from a Single Polarization Observation

     Jue Wang, Ruilin Huang, Liquan Shen et al. · arXiv:2610.25326

     AI summaryMPIG predicts missing 45°, 90°, and 135° polarization views from a single 0° image using a Swin Transformer and Stokes‑aware supervision, achieving 29.7 dB PSNR, 0.9505 SSIM, and 21.5° AoLP error while preserving task‑relevant information.

180. CoRe-WM: Conditioned Response-State World Modeling for Longitudinal Treatment-Response Prediction in Breast DCE-MRI

     Lin Zhao, Bowen Chen, Qiaohui Gao et al. · arXiv:2610.25333

     AI summaryCoRe‑WM models longitudinal breast DCE‑MRI as a state‑transition problem, conditioning the transition on patient profile and therapy, and factorizes response dynamics to lesion morphology. It outperforms all baselines in pCR prediction across multiple decision points and transfers robustly to a different cohort.

181. Edge Deployable Lightweight Framework for Real-Time Driver Safety System with Integrated Lane Departure and Traffic Sign Recognition

     Hassaan Ali Qureshi, Danish Hussain, Ali Hassan et al. · arXiv:2610.25340

     AI summaryUnified edge ADAS that simultaneously performs lane‑departure warning and traffic‑sign detection/recognition, achieving 15 FPS on a Jetson Orin Nano with 97.4 mAP and 95.6 % classification accuracy while keeping the model at 13.9 M parameters.

182. CASE: Complementary Articulation-Aware Speech Encoder to Enhance Speech-Driven 3D Facial Animation

     Jaehong Myung, Seungho Eum, Taegon Kim et al. · arXiv:2610.25342

     AI summaryCASE trains a speech encoder by contrastively aligning audio with 3D lip‑mesh VQ‑VAE latents, then fuses it with a baseline SSL model via cross‑attention. The adapted encoder lowers mean lip‑vertex error by 2.8–6.0% and reduces full‑face error across four VOCASET generators.

183. MCSplat: Multi-View Photometric and Geometric Consistent Feed-Forward Gaussian Splatting for Driving Scenes

     Yifei Wang, Wei Tian, Wenhua Bao et al. · arXiv:2610.25345

     Driving-scene reconstruction via feed-forward Gaussian splatting from sparse multi-view images remains challenging due to two factors. First, cross-camera photometric inconsistency, caused by variations in exposure, ISP, and illumination, can be incorrectly absorbed into…

184. Predicted Class Correction: Conformal Thresholds for Classes with No Calibration Data

     One Octadion, Novanto Yudistira, Yuta Nakashima · arXiv:2610.25349

     Class-conditional conformal prediction calibrates a separate score threshold for every class and therefore requires labelled calibration examples of every class. In long-tailed deployments these are exactly what is missing: rare classes are left with few or no held-out examples,…

185. Local Evidence, Global Failure: Diagnosing Speech-Vision Temporal Composition in Audio-Visual Language Models

     Jaehong Myung, Seunghun Oh, Unsang Park · arXiv:2610.25352

     AI summaryThis paper introduces Temporal Evidence Decomposition, a training-free diagnostic framework that separates local visual verification from speech-relative aggregation in audio-visual language models. It reveals that while models struggle with full-video tasks, their local visual evidence recovery dramatically improves when queried with isolated temporal windows.

186. LateMatch: Breaking the Interaction Bottleneck via Cacheable Matching for Efficient Visual Localization

     Mohammad Omama, Aeyan Ashraf, Sandeep Chinchali · arXiv:2610.25358

     Visual localization is critical for embodied agents, but deploying state-of-the-art deep feature matchers on edge devices is fundamentally limited by their massive computational cost. Standard deep matchers tightly interleave intra-image context building and inter-image matching…

187. Painting Change Map as guidance for generating painting processes

     Markus Pobitzer, Chang Liu, Niccolò Biondi et al. · arXiv:2610.25369

     Digital art generation has undergone a massive paradigm shift, moving from static image synthesis to dynamic video generation that can mimic the human painting process. While existing models can synthesize visually stunning time-lapses of paintings being created or "unpainted,"…

188. Only Means Only: what a harmonization benchmark score averages over

     Imad Buljić, Delaida Muminović · arXiv:2610.25387

     AI summaryThe paper shows that the image‑harmonization benchmark inadvertently includes many test images derived from photographs also used in training, skewing performance scores. The authors release a cleaned test list and report that scores drop significantly when duplicates are excluded.

189. ATLAS: Attention-Guided Test-Time Latent Selection for Compositional Visual Autoregressive Generation

     Hossein Shahabadi, Raya Namazi, Mahdieh Soleymani Baghshah · arXiv:2610.25390

     AI summaryATLAS is a training‑free framework that uses a frozen model’s attention maps to steer test‑time latent selection for compositional image generation. It prunes unpromising paths early, achieving over 10% higher T2I‑CompBench scores while cutting generation time by >50%.

190. A 3D-Driven Framework for Moving-Camera Perspective-to-360{\deg} Video Generation

     Michael Wakeham, Mrinmoy Kumar Das, Wenqi Wang et al. · arXiv:2610.25395

     Turning a moving-camera perspective video into a 360{\deg} video requires more than just generating content outside the input frame: observations revealed over time must be preserved, unseen regions completed, and the resulting scene kept consistent as the camera viewpoint…

191. PolarVis: A Dataset and Diffusion Framework for RGB-to-Polarization Perception in Adverse Driving Scenes

     Ruilin Huang, Jue Wang, Zhen Tian et al. · arXiv:2610.25396

     Reflections, glare, transparent structures, and material-dependent appearance variations can introduce ambiguity into RGB-based visual perception in vehicle-relevant environments. Polarization provides complementary reflection- and material-sensitive information, but its…

192. RAHF-Net: Range-Aware Heterogeneous Fusion for LiDAR-Assisted RGB Semantic Segmentation

     R. D. Lin, Pengcheng Weng, Ge Wang et al. · arXiv:2610.25400

     This paper studies LiDAR-assisted RGB semantic segmentation, where LiDAR point-cloud features improve dense pixel-wise predictions in the RGB image plane rather than 3D point labels or detections. Existing 2D fusion methods often project point clouds into shallow depth/range…

193. MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion

     Houyuan Chen, Fanzhou Wang, Xiangyu Fan et al. · arXiv:2610.25404

     Multi-turn image editing aims to modify an image according to a sequence of instructions, enabling more effective interactive visual content creation for users. However, existing methods often exhibit noticeable drift after only a few editing turns, manifesting in two aspects:…

194. UniState: Unifying Memory and Intervention for Controllable World Simulation

     Jaeseok Jeong, Kyungmook Choi, Sohyun Chung et al. · arXiv:2610.25414

     World models offer increasingly rich environments for exploration, yet intervening in a world while observing the consequences from freely chosen viewpoints remains challenging. We tackle these challenges by introducing UniState, a framework for controllable world simulation…

195. From Label Priors to Task Evidence: Efficient Long-Video Frame Selection via Bayesian GRPO

     Shuochen Chang, Bingjie Gao, Qingyang Liu et al. · arXiv:2610.25418

     AI summaryThis paper presents a Bayesian GRPO framework that transforms static keyframe labels into task‑aligned frame selection for long‑video understanding. It achieves competitive accuracy with heavy MLLM scorers while remaining lightweight, and transfers across downstream MLLMs without retraining.

196. Preserving the Manifold: Frozen Expert Alignment for Label-Efficient SAR Scene Understanding

     Eli A. Underwood, I. Taylor McKechnie, Leslie M. Collins · arXiv:2610.25429

     Cross-modal alignment of synthetic aperture radar (SAR) imagery to natural language is constrained by two distinct failures. The first is modality mismatch: text encoders pre-trained on optical imagery, such as CLIP, cannot be efficiently fine-tuned on SAR, because active…

197. DragonMoCap: Monocular Video-based Motion Capture for Diverse Dragon Topologies

     Zhongfei Qing, Xiaoxiang Zhao, Ling-Hao Chen et al. · arXiv:2610.25437

     This work explores monocular motion capture for dragon-type characters, which currently relies on manual keyframing due to the lack of high-quality datasets and parametric models. The primary obstacle lies in the extreme topological diversity across dragon rigs, where varying…

198. A New Hope for PRNU on Modern Smartphones? Evidence That Device-Specific Fingerprints Survive Computational Photography

     Benjamin Loison (Agence Ministérielle pour l'IA de Défense, France, ENS Paris-Saclay et al. · arXiv:2610.25438

     Photo Response Non-Uniformity (PRNU) is widely regarded as the most reliable marker for source camera identification. Recent studies, however, report high correlations between images and fingerprints from different modern smartphones, which calls into question the discriminative…

199. UBD-Bench: Benchmarking and Analyzing Unimodal Branch Degradation in Multimodal Perception for Autonomous Driving

     Ang Liu, Xiang Gu, Ruikai Li et al. · arXiv:2610.25448

     Multimodal perception improves autonomous-driving performance by combining complementary sensor information. However, stronger multimodal predictions do not necessarily imply stronger representations in each modality-specific branch. Such branch-level weaknesses can remain…

200. BiDex-100: A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks

     Yong Wang, Kecheng Zheng, Xing Zhu et al. · arXiv:2610.25459

     Bimanual manipulation with dexterous hands has attracted growing attention in research on robotic action models. However, few studies have developed benchmarks specifically tailored for dexterous manipulation, leaving the dexterous manipulation capabilities of existing models…

201. Reliability-Preserving Compression for Deployable Medical Image Segmentation

     Sagar Kumar, Shailja Sharma, Tirthajit Baruah et al. · arXiv:2610.25460

     Post-training compression of pretrained medical segmentation models commonly uses quantization, pruning, or distillation to reduce model size and computational cost. Existing medical segmentation quantization studies mainly evaluate Dice, model size, memory, and latency, but…

202. A failure of Imagination: VLMs fail to use visual representations for text, but self-distillation can help

     Shashwat Singh, He He, Tal Linzen · arXiv:2610.25470

     Vision-language (VL) models are typically built by fine-tuning a language model to attend to visual features in service of a Visual Question-Answering objective. We ask whether this alignment changes how language models represent text---especially on tasks where visualization…

203. Test-Time Scaling of Text-to-Image Flow Matching Models with Noise-Rejuvenated Sequential Monte Carlo

     Wenyi Lian, Zheng Zhao, Thomas B. Schön · arXiv:2610.25479

     Test-time scaling of diffusion models with sequential Monte Carlo (SMC) has become a promising strategy for sampling under user-specified rewards, such as human preference scores. However, the standard SMC framework typically relies on stochastic proposals and therefore does not…

204. Multimodal-Conditioned Latent Diffusion Models for Spatiotemporal Brain MRI Forecasting

     Antonio Scardace, Silvia Ingala, Sebastiano Battiato et al. · arXiv:2610.25481

     Spatiotemporal forecasting of 3D anatomical changes is a critical generative vision challenge for modeling Alzheimer's Disease (AD) progression. While brain MRI forecasting from single-modality data has advanced, other key complementary multimodal biomarkers remain…

205. PRIMA: Pose Refined Inductive Motion Adapter for Zero Shot Social Bodily Behavior Recognition

     Mingcong Ding, Babette Bühler, Tim Fütterer et al. · arXiv:2610.25482

     Recognizing fine-grained bodily behaviors, such as self-grooming, postural shifts, and backchanneling, is a useful intermediate step for computational social signal analysis. But the behaviors that matter vary across social contexts, and retraining a model for each new taxonomy…

206. PCXR-Copilot: Device-aware pediatric chest radiographic analysis with segmentation, suppression, and edema-related cues

     Xueyang Li, Alec Girten, Peng Zeng et al. · arXiv:2610.25486

     Patients admitted to pediatric ICU and cardiac ICU often require intensive perioperative support and frequent chest radiography to monitor support devices, extravascular lung water, and body-wall edema. Radiographic interpretation and analysis is often difficult due to…

207. SlimFormer++: Token Mixing Where It Matters for Efficient 3D Medical Image Segmentation

     Yang Hong, Lei Zhang, Zhihong Liu et al. · arXiv:2610.25488

     To determine where computation is needed in volumetric segmentation networks, we analyze block-level redundancy and stage-wise token-mixer contribution. Centered Kernel Alignment reveals high input-output representational similarity within blocks of representative convolutional,…

208. GroundedVDR: Grounding Video-to-Web Evidence Dependencies in Video Deep Research

     Xiyuan Liang, Yu Wang, Zuhao Yang et al. · arXiv:2610.25493

     AI summaryGroundedVDR benchmarks video‑to‑web evidence grounding by requiring human‑verified video segments and web evidence annotations and scoring each component separately. The best multimodal agent achieves 76% answer accuracy, but only 8% of questions retrieve all required web evidence.

209. From Seeing to Reporting: A Theory of Reliable Visual Set Generation in Vision-Language Models

     Junzhe Shi, Yuan Gan, Shida Jiang · arXiv:2610.25503

     Vision-language models (VLMs) are asked to report what they see: to count objects, box them, parse a screen or pick a click target (GUI grounding). We argue that reporting is part of the visual task, not merely its formatting. In a dense scene with 173 shapes, GPT-6.1 lists 173…

210. ReDT: Resolution-Decoupled Visual Geometry Transformer

     Abteen Arab, Guile Wu, Dongfeng Bai · arXiv:2610.25517

     We introduce ReDT, a feed-forward visual geometry transformer designed to scale multi-view reconstruction to high-definition (HD) inputs. Existing visual geometry transformers are typically trained at fixed, standard definition (SD) resolutions and exhibit degraded…

211. Modaliff: A Unified Token Diffusion Model for Any-to-Any Generation

     Duc Kieu, Kien Do, Minh Khoa Le et al. · arXiv:2610.25533

     AI summaryModaliff learns a token-based diffusion model that uses bimodal pairs to keep each modality’s native latent sequence while unifying them into a single token stream with per‑token attributes. It outperforms other any‑to‑any models in fidelity across text, image, and audio with fewer parameters and training data.

212. Residential Construction Image Dataset Using CSI Divisions for Deep Learning-based Object Detection

     Fahad Iqbal, Shayan Mirzabeigi · arXiv:2610.25535

     Accurate recognition of construction site elements is essential for the application of deep learning (DL) in the construction industry. Despite rapid advances in computer vision, the adoption of DL in construction has been limited by the lack of structured, publicly available…

213. StreamAhead: Real-Time Streaming Video Question Answering via Adaptive Precomputation

     Haoyang Fan, Dhruv Parikh, Rajgopal Kannan et al. · arXiv:2610.25551

     Real-time streaming video question answering with vision-language models must answer queries that arrive at any time with low latency while sustaining the video-processing rate required by the application. Existing systems construct key-value (KV) caches for the entire video as…

214. Tricking Synthetic Image Detectors with Amortized Generative Attacks

     Mohamed Awad, Khaled Elhady, Yasser Rohaim et al. · arXiv:2610.25554

     Rapid advances in image synthesis have increased the need for reliable synthetic-image detectors, yet their robustness to transferable attacks remains incomplete and understudied. Existing evaluations generally examine whether an attack constructed against one detector transfers…

215. DiffGround: Semantic Difference-Guided Zero-Shot Video Question Grounding

     Cheng-Han Lee, Amir Salimi, Hsiang-Wei Huang et al. · arXiv:2610.25570

     AI summaryDiffGround uses a frozen multimodal LLM to extract an action cue from a question, turning it into a retrieval query that guides a VLM to pick keyframes. Semantic-difference refinement recovers full action segments, achieving state‑of‑the‑art zero‑shot video question grounding.

216. Finite Interaction Lens: Structured Precision Sharing for Vision-Language Models

     Nozomu Fujisawa · arXiv:2610.25577

     Identifying which visual interventions interact in a vision-language model requires comparing their joint and separate effects. The resulting four-output difference can be small enough for rounding error to obscure it. We introduce the Finite Interaction Lens (FIL), a…

217. ViMoLa: Discrete Motion Tokens for Language-Guided Video Generation

     Bishoy Galoaa, Sarah Ostadabbas · arXiv:2610.25595

     Language-guided video generation often captures scene content while missing the requested motion. We present ViMoLa (Video Motion Language), which predicts a motion prior from language and injects it into a frozen image-to-video backbone. A spatial vector-quantized autoencoder…

218. CheXFuse: Guided Differential Joint Attention Fusion for Robust Multimodal Chest X-Ray Classification in Data-Scarce Medical Settings

     Muntaqim Ahmed Raju, Priyanka Siddappa, Md Shifat Haider Al Amin et al. · arXiv:2610.25617

     A central challenge in multimodal learning is how to fuse modalities that differ substantially in predictive strength. When one modality provides a much stronger representation than another, indiscriminate cross-modal interaction can introduce weak, redundant, or unstable…

219. Recurrent Vision Transformers for Event-Based Spacecraft Pose Estimation Under Slow Relative Motions

     Sofia McLeod, Mohsi Jawaid, Carl Seubert et al. · arXiv:2610.25622

     Vision-based object pose estimation is vital for conducting spacecraft rendezvous and proximity operations (RPO). Recent research has demonstrated the robustness of event cameras against the harsh lighting conditions that often occur in orbit, thanks to the extremely high…

220. Where Do the Gaussians Go? Foreground Attraction for Camera-Based Weakly Supervised 3D Occupancy

     Jeagon Song, Jhonghyun An · arXiv:2610.25631

     Camera-based 3D semantic occupancy models can be trained without 3D labels by rendering their predictions into the images and supervising them with 2D pseudo labels. In sparse Gaussian representations trained this way, Gaussian positions are learned only through pixel-wise…

221. Confidence-Guided Pseudo-Label Refinement for Unsupervised Cross-Domain Person Re-Identification

     Sundas Iqbal, Danish Ali, Qing Tian et al. · arXiv:2610.25633

     Unsupervised cross-domain person re-identification remains challenging because domain shifts generate noisy pseudo labels that are reinforced during iterative adaptation. Existing clustering-based methods often rely on limited reliability cues or global confidence criteria and…

222. Clean Forcing: Drift-Resistant Autoregressive Video Diffusion with a Frozen Base

     Wenqing Wang, Joonghyuk Shin, Jonathan Tremblay et al. · arXiv:2610.25634

     Autoregressive (AR) video diffusion models generate streaming video of arbitrary length by conditioning each new chunk of frames on those already generated. However, they suffer from exposure bias, where a model trained on clean context must continue from its own imperfect…

223. TNVD-SC: Diagnosing Scan Context Failures in Forests with Terrain-Normalized Vertical Distributions

     Jineon Kim, Jhonghyun An · arXiv:2610.25637

     Long-term LiDAR place recognition in forests is difficult because repetitive vegetation and uneven terrain create severe structural aliasing. We diagnose Scan Context (SC) by separating candidate retrieval from fine matching: on the standardized TreeLoc++ Wild-Places protocol, a…

224. ACorD: Adaptive Correspondence Discovery for Medical Image Registration

     Arghavan Rezvani, Kun Han, Xiaohui Xie · arXiv:2610.25638

     Medical image registration is a fundamental task for aligning anatomical structures across images. Exhaustively searching for point-wise correspondences is prohibitively expensive in both computation and memory, particularly for high-resolution 3D medical images. Most recent…

225. Visual Memory and Textual Alignment for Cross-Modal Few-Shot Segmentation

     Tianqi Wei, Zhi Chen, Xin Yu et al. · arXiv:2610.25641

     AI summaryViTA‑SAM augments SAM3 with a visual memory that stores support masks and a cross‑modal alignment module that conditions textual features on this memory, plus pseudo‑label refinement. It achieves >5% mIoU gain over prior methods on FSS benchmarks.

226. Negation without Confusion: Decomposing Bias and Recomposing Style in Text Embeddings for Negation-aware Image Generation

     Sooyeon Park, Hemraj Singh, Kyunghyun Min et al. · arXiv:2610.25645

     Negation-aware image generation remains challenging for text-to-image diffusion models, which often struggle with negation due to implicit linguistic biases that are deeply entangled within text representations. As a result, bias-aligned semantics frequently dominate over…

227. InteractAgent: Agentic Human Motion Interaction with Memory-Augmented Reflection

     Tao Zhang, Senhe Zhang, Zeyu Zhang et al. · arXiv:2610.25649

     Long-horizon human-scene interaction requires coordinated spatial reasoning, physical feasibility, and the ability to recover from compounding execution errors. Text-driven motion models are largely scene-agnostic and lack closed-loop correction, while geometry-conditioned…

228. FluidGen: Physics-Aware Fluid Video Generation via Motion-Guided Post-Training

     Yang Zheng, Zizhang Li, Boyang Deng et al. · arXiv:2610.25660

     AI summaryFluidGen augments a pre-trained video generator with a dual‑stream that predicts velocity fields from high‑fidelity Houdini simulations, regularizing outputs toward physically consistent fluid motion. The resulting model outperforms baselines, producing realistic smoke, fire, and debris dynamics.

229. PoSTCache: From Forecasting to Postdictive Smoothing for Revisable Diffusion Caching

     Liwei Lan, Hui Chen, Jungong Han et al. · arXiv:2610.25664

     Diffusion and flow matching models achieve impressive visual generation quality but remain computationally expensive due to iterative sampling. Caching-based acceleration provides a practical training-free solution by reusing or predicting intermediate features. However,…

230. SpaRV: Sparse Radiance Voxels for Scalable Geometrically Consistent LiDAR-Visual Mapping and Rendering

     Jianheng Liu, Yunfei Wan, Fangming Cheng et al. · arXiv:2610.25665

     AI summarySpaRV builds a sparse radiance voxel map by jointly optimizing geometry with LiDAR depth and images, alternating with camera pose/focal length refinement. It keeps geometry consistent beyond captured trajectories and produces photorealistic, scalable maps on a single 24 GB GPU.

231. Frechet Forcing: Training Long-Horizon Streaming World Models from Short Clips

     Yian Zhao, Yu Yan, Hongcan Guo et al. · arXiv:2610.25666

     Streaming video world models generate the future chunk by chunk, conditioned on their own past and a stream of actions. Trained on short clips, they drift once a rollout exceeds the training horizon, as prediction errors re-enter the context and compound. A common remedy is to…

232. RxGS: Relightable Gaussian Stylization via BRDF Decomposition

     Wanrong Hong, Liu Jiang, Mukesh Prasad et al. · arXiv:2610.25681

     Reference-based 3D scene stylization transfers the appearance of a reference image onto a reconstructed 3D scene while preserving scene geometry. Recent methods perform this transfer by optimizing the appearance representation of neural scene models including Neural Radiance…

233. BitFlat: Error-Localized Bitstream Design for Robust Diffusion Image Compression

     Xinfeng Deng, Li Zhou, Haitao Chen et al. · arXiv:2610.25682

     Zero-shot diffusion image compression based on reverse channel coding achieves high perceptual quality at low bitrates without retraining, yet its bitstream remains vulnerable to bit flips. Prior work has shown that a joint combinatorial representation can amplify a single bit…

234. Multi-View Structural Representations for Physically-Consistent TeX Generation

     Daoheng Li, Jianjun Qian*, Mengkai Yan et al. · arXiv:2610.25695

     Infrared and visible images provide complementary cues for low-light image generation, and physics-based methods can better exploit this complementarity. However, existing physics-based methods still rely heavily on paired data. To address these limitations, we propose…

235. LEVI: Benchmarking Visual Identity Fidelity with Verifiable Low-Entropy Constructs

     Xiao Ma, Zhonghao Zhan, Hamed Haddadi et al. · arXiv:2610.25716

     Modern generative models can produce visually plausible images while violating exact identity constraints. Evaluating such failures requires distinguishing violations of the requested identity from permissible variation in appearance. We introduce Low-Entropy Visual Identity…

236. Think Densely, Act Sparsely: Latent Expert Cognitive Chains for Vision-Language-Action Autonomous Driving

     Jie Wang, Guang Li, Zhijian Huang et al. · arXiv:2610.25734

     Autonomous driving fundamentally demands dense understanding of visual, semantic, and geometric information, yet existing Vision-Language-Action (VLA) models are typically trained with sparse supervision such as language instructions or trajectory signals. This direct mapping…

237. Functional Cross-layer Experts Routing for One-step Diffusion Distillation

     Zile Huang, Ser-nam Lim · arXiv:2610.25742

     One-step distillation compresses diffusion trajectories into a single forward evaluation, yet how internal capacity should be structured within that remaining pass remains largely unexplored. In this setting, identical parameters must simultaneously preserve general generative…

238. MAPS: A Multi-Agent Personalized Script-Driven Video Advertising System Enhanced by Reinforcement Learning

     Zhongyu Wang, Chen Zuo, Qingxuan Sun et al. · arXiv:2610.25748

     With the rapid growth of online video consumption, video advertising has become a dominant component of digital advertising. Most personalized video advertising systems follow a retrieval-based paradigm, which matches ad content with user preferences in a feature space and…

239. VHA: VQA-based Hallucination Assessment for Generative Image Super-Resolution

     Donghun Ryou, Inju Ha, Jaeyoo Park et al. · arXiv:2610.25770

     Generative image super-resolution produces visually realistic high-resolution images, but may also introduce hallucinations by adding unsupported content or altering existing details. Because existing super-resolution metrics fail to measure whether generated details remain…

240. See the Future, Classify the Present: Stochastic Latent World Modeling for Hyperspectral Images

     Muhammad Ahmad, Gemine Vivone, Ronny Hänsch et al. · arXiv:2610.25781

     Hyperspectral image classification (HSIC) relies on static spatial-spectral representations learned from labeled samples, limiting its ability to model structured dependencies in the latent representation space. To address this limitation, this work proposes Stochastic Latent…

241. UniJev: Separating Accuracy from Calibration in Trimodal System One Decision Models

     Jianing Wang, Youwen Jian, Derong Wang et al. · arXiv:2610.25784

     System One models produce probabilistic judgments that software can consume directly, making reliable confidence as important as answer accuracy. Yet joint training across text, images, and video remains underexplored, and the contributions of supervised fine-tuning (SFT) and…

242. Do We Need Learned Targets? Classical Gradients as Supervision for 2D Gaussian Initialization

     Kyungjune Baek · arXiv:2610.25809

     Learning-based Gaussian image initializers can require costly supervision obtained by per-image optimization. We investigate whether the Pseudo-PPM targets used by Instant GaussianImage can be replaced by classical image gradients. Our framework uses normalized Sobel maps in two…

243. Hihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing

     Zheng Chen, Yuchen Huang, Xiaoyang Liu et al. · arXiv:2610.25832

     Diffusion models have recently achieved encouraging results in real-world image dehazing. However, restoring high-resolution images remains challenging due to spatially non-uniform haze and inconsistent predictions across local patches. To address these issues, we propose…

244. SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation

     Yanjin He, Cuifeng Shen, Wanzhou Lei et al. · arXiv:2610.25844

     Training high-quality diffusion models is computationally expensive, and inference-time guidance often adds further sampling cost. In this work, we introduce Swift, a simple yet effective self-guided training method that uses a frozen weak denoiser to improve a stronger model…

245. Auditing Semantic and Relational Reliability in Open-Vocabulary Human-Object Interaction Evaluation

     S M Asif Hossain, Ruksat Khan Shayoni, Md Kishor Morol · arXiv:2610.25849

     Open-vocabulary human-object interaction detectors describe interactions in free text, and this creates a problem for evaluation. Category scoring rejects valid paraphrases, while semantic similarity scoring can reward predictions that name the wrong action or reverse the roles…

246. ActionRoPE: action as coordinates for spatially consistent control in planar-camera game world models

     Zhiyang Deng, Yuzhe Wu, Zijun Lin et al. · arXiv:2610.25861

     Game world models have made rapid progress toward real-time interactive generation. Yet action control is still predominantly implemented as feature conditioning, leaving the model to learn from training data how each action should move the visual world. This formulation is…

247. Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation

     Lingsen Zhang, Wei Li, Jie He et al. · arXiv:2610.25874

     Robotic manipulation is not merely observation-to-action generation, but the realization of task-specified state changes in the physical world. Existing vision-language-action (VLA) methods directly learn observation-to-action mappings, imposing a substantial learning burden due…

248. Rethinking Sparse-View 3D Gaussian Splatting: A Few-Shot Learning Perspective on Overfitting Prevention

     Lejia Ye, Bin Wu, Jingjing Gao et al. · arXiv:2610.25887

     Sparse-view novel view synthesis (NVS) is essential for AR/VR and real-time communication applications, yet remains challenging when reconstructing 3D scenes from limited viewpoints. While 3D Gaussian Splatting (3DGS) excels in dense-view scenarios, it suffers from overfitting…

249. Out-of-Distribution Detection in Continual Learning

     Nimeshika Udayangani, Flora D. Salim, Christopher Leckie et al. · arXiv:2610.25901

     Out-of-distribution detection in continual learning (OOD-CL) is highly challenging, as it requires (i) learning a sequence of tasks without catastrophic forgetting, and (ii) identifying unknown samples (OOD) without access to known (in-distribution (ID)) data from past tasks,…

250. Multi-scale Fusion of Fast Vision Mamba and Foundation Model Features for Patch-Level Gleason Grading in Computational Pathology

     Zihe Zhang, Weimiao Yu, Piumi Don Simonge · arXiv:2610.25904

     Automated Gleason grading of prostate whole slide images (WSIs) enables faster and more consistent cancer risk stratification. However, it remains a challenging task due to the subtle morphological differences between grades, the severe class imbalance arising from rare…

251. LEAP: Accelerating ODE-based Diffusion Sampling with Optimal Linear Leaps

     Haowei Zheng, Zhenyu Zhou, Jiawei Chen et al. · arXiv:2610.25918

     Diffusion models primarily rely on general-purpose ODE-based numerical methods for training-free accelerated sampling, but performance degrades with coarse time discretization. In this paper, we propose optimal linear leaps (LEAP), a specialized method grounded in theoretical…

252. SAPO: Self-Supervised Attention Policy Optimization for Improving Temporal Video Grounding in Vision-Language Models

     Zixuan Chen, Depeng Wang, Wei Chen et al. · arXiv:2610.25923

     Video vision-language models (VLMs) have strong semantic understanding, yet remain unreliable at precise temporal grounding. Existing gains from explicitly aligning video frames with absolute time suggest that VLMs may already encode accurate temporal locations internally but…

253. MIRA: A Multimodal Illustrated Deep Research Agent for Controllable Reports

     Hongrui Wu, Kaiwen Tuo, Congcong Wang et al. · arXiv:2610.25926

     AI summaryMIRA is a deep research agent that integrates text and images across search, writing, revision, and delivery stages, using a generation contract to control report granularity, multimodal composition, and evidence grounding. It outperforms existing open agents on seven multimodal-search and deep-research benchmarks while producing traceable illustrated reports.

254. CondenseVLA: Learnable History Condensation for Efficient Multi-Frame VLA

     Feiyang Hong, Yujie Wei, Bo Zhao et al. · arXiv:2610.25927

     Robot manipulation often depends on short-term cues - such as brief contacts, temporary occlusions, and intermediate object states - that are ambiguous in a single frame and can only be resolved by integrating observations over time. However, injecting historical information…

255. OmniGround: Strengthening Omni-Modal Temporal Grounding via Attention-Consistency Training

     Zixuan Chen, Depeng Wang, Hao Lin et al. · arXiv:2610.25929

     AI summaryOmniGround improves temporal grounding in omni‑modal models by aligning internal attention with predictions and ground truth through Temporal Attention‑Consistency Training (TACG). It achieves state‑of‑the‑art mIoU (37.5 on LongVALE, 35.7 on UnAV‑100) without extra temporal cues.

256. HORIA: 4D Human and Object Reconstruction from a Monocular Video with Interaction-Aware Re-Rendering

     Hyeongjin Nam, Taeryung Lee, Daniel Sungho Jung et al. · arXiv:2610.25960

     Joint 4D reconstruction of a human and object from a monocular video is essential for understanding how people interact with their environment. While recent methods have made noticeable progress, capturing such human-object interactions remains difficult due to two main…

257. Token-Set Choice Confounds POPE: A Systematic Audit of Yes/No Extraction in VLM Hallucination Evaluation

     Kesav Kumar Jayakumar, Karthigeyan Thilak · arXiv:2610.25964

     We set out to reproduce a recent hallucination-mitigation result on the POPE benchmark. The LLaVA-1.5-7B baseline we measured, 82.21 F1 on the adversarial split, sat 1.21 to 8.67 points above the adversarial baselines we planned to compare against (81.00, 79.30, and 73.54). We…

258. On the Benefits of Uncertainty Estimation for Species Classification: A Case Study on Optimizing Expert re-Annotation

     Kamenan N'Gadi (LORIA, BISCUIT), Martin Laviale (LIEC et al. · arXiv:2610.25977

     Automated identification of diatoms from microscopic images is a cornerstone of large-scale freshwater biomonitoring, yet the high diversity of species and strong class imbalance pose significant challenges for reliable classification. In this study, we develop a ResNet-18…

259. TempoMoE: Identifying and Steering Fast-Slow Thinking Experts for Efficient Multimodal Reasoning

     Xiyue Zhang, Keru Zhou, Ziyi Wang et al. · arXiv:2610.25999

     Multimodal Mixture-of-Experts (MoE) reasoning models rely on long chain-of-thought (CoT) for complex tasks, yet often overthink easy inputs, increasing token cost and latency, and sometimes degrading accuracy. Existing methods optimize reasoning length during training or impose…

260. Curvature as a Baseline: Monocular Pose Measurement of Cylindrical Objects with Dual Colour Stripe Markers

     Jaesung Hong, Dong Won Lee · arXiv:2610.26003

     Thin cylindrical objects offer little space for planar markers. We propose a dual-colour-stripe marker that uses the known curvature of an object to improve pose measurement with a single camera. The boundary between two colours provides a long centre line, while dots and…

261. FrameGRPO: Incentivizing Video Reasoning via Multi-Frame Rollouts and Temporal Credit

     Jiaxuan Sun, Biao Wang, Shuai Li · arXiv:2610.26022

     AI summaryFrameGRPO improves video reasoning for multimodal LLMs by using multi‑frame rollouts that sample different temporal views and a gated temporal credit that rewards short‑frame trajectories only when full‑frame references fail, yielding better performance on video benchmarks.

262. TamedFlow: Trajectory-Aware Distillation\\for Teacher-Assisted Diffusion Sampling

     Shuaiqi Cheng, Jiahao Zhang, Boyuan Gu et al. · arXiv:2610.26025

     Diffusion Transformers (DiTs) deliver high-fidelity image and video generation, but inference requires many sequential model evaluations. Two acceleration strategies are increasingly important: model compression makes each evaluation cheaper, while adaptive inference avoids…

263. After the Federation: Benchmarking Knowledge Portability to Heterogeneous Newcomers

     Xinyuan Zhao · arXiv:2610.26026

     Once historical clients become unavailable, a federation must serve newcomers through the artifact it leaves behind. How much of its knowledge remains useful in an architecture chosen afterward? We introduce FEDNBO, a seven-dataset benchmark that separates knowledge supply,…

264. AVTime: Reinforcing Dense Audio-Visual Captions with Bidirectional Alignment

     Mingyang Wu, Kaituo Feng, Bohao Li et al. · arXiv:2610.26046

     AI summaryAVTime introduces a 50K‑scale audio‑visual caption dataset with fine temporal resolution and bidirectional alignment, a model trained with a Bidirectional Temporal‑Semantic Alignment reward, and a benchmark evaluating both semantic correctness and temporal accuracy.

265. Procedural Exposure Bias in Skill-Augmented Language Model Agents

     Qing Yang, Yuqi Xiong, Pengcheng Huang et al. · arXiv:2610.26062

     External skills provide language-model agents with reusable procedural knowledge, yet how models use such knowledge after it is supplied remains poorly understood. We identify Procedural Exposure Bias (PEB), the tendency of language models to increasingly favor an action as its…

266. Learning the Dose Manifold: A Latent Diffusion Framework for Volumetric Dose Prediction in Automated Radiotherapy Treatment Planning

     J. Boyer-Chammard, R. Vauclin, M. C. Biston et al. · arXiv:2610.26064

     Automated treatment planning (ATP) aims to reduce the unwarranted inter-planner variability of manual planning, while improving treatment quality and efficiency. Most ATP pipelines rely on the prediction of the achievable three-dimensional dose distribution for a given anatomy…

267. PhysSkin++: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning

     Yuanhang Lei, Baofei Liu, Xingxuan Li et al. · arXiv:2610.26067

     Generalizing real-time physics-based animation across diverse 3D shapes and spatial discretizations remains challenging. In this paper, we introduce PhysSkin++, a physics-informed framework that learns continuous neural skinning fields from static geometries to construct compact…

268. Risk Decomposition for Hallucination-Aware Video Sampling

     Chenyan Zheng, Fan Wu, Ruibo Li et al. · arXiv:2610.26069

     Despite achieving high visual fidelity, video models still suffer from physical hallucinations, limiting their reliability as world simulators. Existing methods leverage model-internal or external signals to improve physical plausibility, but their roles and limitations require…

269. Not All Similarity Is Leakage: Distribution-Aware Identity Leakage Assessment in Synthetic Face Datasets

     Pawe{\l} Borsukiewicz, Daniele Lunghi, Wendkûuni C. Ouédraogo et al. · arXiv:2610.26077

     Synthetic data provides the means for more ethical and less biased facial recognition. Nonetheless, the limitations of generative models and the conditioning process they rely on pose a risk of identity leakage, resulting in synthetic samples that are overly similar to those…

270. DCR-TSF: Dual-Level Compositional Reasoning with Triadic Spherical Fusion for Training-Free Zero-Shot Composed Image Retrieval

     Gwang-Ho Na, Ho-Joong Kim, Seong-Whan Lee · arXiv:2610.26078

     Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image that reflects an intended textual modification while maintaining semantic consistency with a reference image. Existing training-free ZS-CIR methods typically rely on fine-grained generated target…

271. Toolkit for Facial Deidentification

     Esteban Leiva-Montenegro, Matthieu Pillot, Manfred Gonzalez-Hernandez et al. · arXiv:2610.26079

     This work presents a novel open-source toolkit designed for the evaluation of face deidentification techniques with available evaluation models. The framework provides a complete evaluation pipeline, including preprocessing, deidentification, and performance assessment, based on…

272. DeOverlay: A Benchmark and Source-Anchored Flow for Video Overlay Removal

     Deshui Miao, Yuhang Ma, Shaoyong Jia et al. · arXiv:2610.26081

     Video overlay removal aims to recover scenes obscured by subtitles, titles, and logos while preserving surrounding content. Existing methods use temporal propagation or source-conditioned generation, but can leave residual overlays and distort scene structure when useful source…

273. Context-Aware Topological Distillation Graph Network for Asymmetric EEG-Eye Movement Emotion Recognition

     Peijun Shi, Xiujin Zhu, Zhenguo Lao et al. · arXiv:2610.26084

     EEG-eye movement emotion recognition can support mental-health monitoring and adaptive interfaces, but most fusion models treat the two signals as parallel features, classify each window alone and degrade when eye data are lost. We propose the Context-Aware Topological…

274. RGB--Event Object Detection Across Operating Frequencies via Reliability-Supervised Routing and Asymmetric-Resolution Fusion

     Peijun Shi, Jiannan Zhao, Xiujin Zhu et al. · arXiv:2610.26090

     Autonomous vehicles must detect objects faster than a frame camera delivers images. Event cameras record brightness changes with microsecond resolution, so RGB--event detectors can decide between frames, at an operating frequency, the rate of decisions, far above the frame rate.…

275. GeoNaviDiff: Transferring Disentangled Semantic Controls to Diffusion Features via Local Riemannian Navigation

     Tao Sun, Huanting Guo, Baao Xie et al. · arXiv:2610.26091

     Disentangled real-image manipulation requires factor-specific semantic control while preserving image fidelity and non-target content. StyleSpace $S$ provides interpretable controls but is limited by GAN inversion, whereas diffusion models offer stronger image realization…

276. How Robust Are Prognostic Radiomic Features? Comparing Selection Techniques on Pooled Glioblastoma Cohorts

     Yannick Sodonougbo, Marlene Schnieders, Alexia Rizoudis et al. · arXiv:2610.26095

     Radiomic survival studies in glioblastoma report heterogeneous prognostic features, but how much of this stems from the methodology itself is rarely examined. We analysed the association of 463 radiomic, anatomical and clinical features from pre-treatment MRI with overall…

277. PnP-Transformer: A Neural Pose Solver for 6D Object Pose Estimation

     Emirkan Burak Yilmaz, Yakup Genc · arXiv:2610.26096

     We present PnP-Transformer, a neural pose solver designed as an iterative-free, differentiable alternative to classical PnP-RANSAC algorithms. PnP-Transformer regresses the 6D object pose by processing predicted semi-dense 2D-3D correspondence fields alongside confidence scores,…

278. Optimal Multi-Kernel Combination for Online Multi-Label Classification

     Tingting Zhai, Wei Liu, Renjian Yu · arXiv:2610.26123

     Multi-label classification is particularly challenging in streaming scenarios, where data arrives sequentially. Online kernel methods offer a promising solution. However, prevailing methods rely on a single preselected kernel, making them prone to suboptimal choices since the…

279. TriTD: Tri-Partite Trajectory-Distribution Distillation for Real-Time Autoregressive Video Generation

     Jia Li, Xurui Peng, Haowei Zhu et al. · arXiv:2610.26165

     Autoregressive (AR) video diffusion has emerged as a leading paradigm for efficient video generation, yet its practical deployment remains bottlenecked by the trade-off between generation quality and inference efficiency. We attribute this bottleneck to three error sources in AR…

280. Learning Goal-Conditioned Subgoals through Frozen Hierarchical World Models

     Xiang Kaibo, Liu Qingfeng · arXiv:2610.26179

     AI summaryA goal‑conditioned network replaces expensive high‑level planning in hierarchical world models by providing conditioning vectors while a frozen high‑level predictor generates latent waypoints. It matches or exceeds search‑based baselines on several benchmark tasks and speeds up inference 1.3–4.4×.

281. HP-PnP-Flow: Hybrid Time Scheduling and Projection for Plug-and-Play Image Restoration via Flow Matching

     Yong Li, Xiaojie Wang, Xiaoyan Zhang · arXiv:2610.26183

     Image restoration is a fundamental inverse problem that seeks to recover high-quality images from degraded observations. Plug-and-Play (PnP) methods provide a flexible framework for integrating pretrained denoisers into iterative reconstruction, whereas flow matching offers an…

282. Wan-Alpha++: Animating a Single RGBA Asset into Controllable Transparent Video Effects

     Haotian Dong, Xin Wang, Wenjing Wang et al. · arXiv:2610.26189

     RGBA video generation is essential for creating compositable visual assets in filmmaking, game development, and digital content production. However, generating high-quality RGB appearance alongside temporally coherent transparency remains a significant challenge. Existing…

283. GeoEdit: Adversarial De-Evidencing for Silent Geoprivacy Protection via Image Editing

     Yu Yan, Shengjia Cheng, Fei Ma et al. · arXiv:2610.26194

     Recent efforts have demonstrated the strong capabilities of Multimodal Large Reasoning Models (MLRMs) in geographic reasoning. By jointly analyzing subtle visual cues, these models can infer the locations of photos, thus enabling doxing even by non-expert adversaries. Existing…

284. Focused Omni-Modal GUI Agents

     Fuyao Zhang, Wenwen He, Jiaming Zhang et al. · arXiv:2610.26203

     Graphical User Interface (GUI) agents increasingly act over streaming omni-modal observations spanning screenshots, audio, video, interaction history, and device states. However, exposing this complete context uniformly at every step ignores that relevance is transient and…

285. ExMR: Extremity Motion Refinement for Monocular Human Motion Capture

     Tom Wehrbein, Bodo Rosenhahn · arXiv:2610.26213

     Detailed hand and foot articulation is important for animation and motion analysis, yet remains challenging to recover from monocular video. Recent expressive video methods infer hand articulation from shared whole-body representations. Our conference work suggests that this…

286. RoAR-VAE: Improving Latent Diffusability through Robust Prediction

     Yang Yue, Fangyun Wei, Yitong Wang et al. · arXiv:2610.26216

     An effective visual tokenizer must preserve image fidelity while producing latents that downstream diffusion models can learn efficiently. Prior work improves latent diffusability through semantic alignment or joint tokenizer--diffusion training. Revisiting REPA-E, a…

287. SAR2Net: Learning Spatially Anchored Representations for Retrieval-Guided Cross-Stain Alignment

     Tianle Shen, Fang Yan, Xiaofan Zhang · arXiv:2610.26217

     Achieving spatial alignment between whole-slide images(WSIs) across stains is essential for integrating morphology and molecular information, yet it remains challenging due to extreme resolution, tissue fragmentation, and large nonlinear deformations. Conventional registration…

288. RECAP: Reciprocal Sparse Attention and Hierarchical Caching for Efficient Video Diffusion

     Tiange Xu, Kai Liu, Linghe Kong et al. · arXiv:2610.26244

     AI summaryRECAP couples sparse attention with hierarchical caching, using reciprocal correction guided by online error feedback to share refinement information across denoising steps, achieving 3.05× speedup on HunyuanVideo 720p 5 s while improving fidelity over existing sparse‑attention and caching methods.

289. XGLASS: X-Ray, Gaze and Language Assisted Slice Sequencing for CT Volumes

     Michael Glöckler, Timo Ropinski · arXiv:2610.26255

     Vision-language models (VLMs) achieve remarkable performance on 2D medical images, such as X-rays or pathology images, while 3D VLMs that directly process CT/MRI volumes lag behind due to limited 3D training data and the predominance of 2D foundation architectures. To leverage…

290. Dynamics-Aware Sparse Attention for Efficient Autoregressive Video Diffusion

     Yuanyu He, Zhuokun Chen, Yefei He et al. · arXiv:2610.26262

     Autoregressive Diffusion Transformers have emerged as a powerful paradigm for long video generation. However, they are often constrained by prohibitive computational overhead and substantial memory footprints. This inefficiency arises from dense computation mechanisms that…

291. TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation

     Yiheng Dong, Yi Lin, Shilong Huang et al. · arXiv:2610.26271

     AI summaryTIM decouples spatial pathology and temporal evolution in longitudinal radiology report generation, then iteratively refines prior and current reports via mutual feedback, achieving new state‑of‑the‑art language and clinical metrics on Longitudinal‑MIMIC.

292. From Cause to Effect: Generating Intermediate Event Chains in Videos

     Ridouane Ghermi, Vicky Kalogeiton, Ivan Laptev · arXiv:2610.26275

     Video-Language Models (VLMs) excel at visual recognition but struggle with causal reasoning. Current evaluations are flawed: multiple-choice QA allows shortcuts, and open-ended QA lacks robust automatic evaluation. We introduce Causal Chain Generation (CCG), a novel task where…

293. CARNet: End-to-End Coupled Association-Recognition Network for Fine-Grained Multi-Drone Collaborative Detection

     Zhao Zhang, Ruigang Fu, Yu Zhang et al. · arXiv:2610.26312

     Multi-drone collaborative detection aims to enhance target recognition by leveraging complementary observations from overlapping aerial viewpoints. Existing approaches typically follow a two-stage pipeline that first establishes cross-view correspondences and then performs…

294. MCDA: Multi-View Graph Contrastive Learning with Dynamic Topology Augmentation for Functional MRI Analysis

     Hao Zhang, Xiaoyun Liu, Shuo Huang et al. · arXiv:2610.26322

     Resting-state fMRI (rs-fMRI) is a vital non-invasive tool for diagnosing brain disorders and uncovering pathological mechanisms. However, existing graph contrastive learning (GCL) methods often underexploit augmentation histories across training epochs and the relationships…

295. TriScore: Post-Hoc Out-of-Distribution Detection with Energy, Boundary Probes, and Transform Consistency

     Ege Gursoy, Nicolas Mansard · arXiv:2610.26323

     Post-hoc out-of-distribution detection usually relies on a single signal extracted from a frozen classifier, such as confidence, energy, features, or gradients. This dependence becomes fragile under classifier-level adversarial shift, where the shift can preserve the detector's…

296. HGASurv: Hierarchical Granularity-Aware Integration for Multimodal Survival Prediction

     Peike Yu, Wuchao Liu, Jihong Guan et al. · arXiv:2610.26357

     Objective: Multimodal survival prediction jointly analyzes histopathological whole-slide images (WSIs), genomic profiles, and clinical variables to stratify patients by risk. However, existing approaches often inadequately account for granularity differences across modalities,…

297. From Legacy to Contemporary: Scalable Color Modernization for Long Video

     Dongyun Kim, Jiwoo Choi, Jaeseong Lee et al. · arXiv:2610.26391

     Bringing legacy video back to contemporary screens does not necessarily make it look contemporary. Earlier imaging pipelines often leave color casts, muted chroma, weak con- trast, and tonal imbalance, while correcting these character- istics still relies heavily on expert…

298. One Fact, Many Places: Redundant Evidence in Long Visual Documents Breaks Single-Gold Evaluation of Retrieval, Citation and Attribution

     Xiaokang Yuan, Binhui Yao, Luwei Yang et al. · arXiv:2610.26401

     Document understanding is increasingly judged on where an answer comes from, scored against annotated gold locations as if each fact lived in one place. In long visual documents, a fact often lives in several places: a chart value is restated in the text, a key figure reappears…

299. LaMER: Language-Conditioned Multi-Expert Routing for Weakly Supervised Referring Expression Comprehension

     Yu Yu Chang, Tzu-Hsuan Weng, Youssef Mohamed Amin et al. · arXiv:2610.26404

     Referring Expression Comprehension (REC) localizes the image region described by a natural language expression, underpinning applications such as autonomous navigation, human-robot interaction, and medical image analysis. Under weak supervision, REC is challenged by the…

300. Policy-Adaptive Credit Assignment for Spatial Reasoning with Verifiable Rewards

     Hailiang Yang, Shengchao Chen · arXiv:2610.26410

     Image-based spatial reasoning requires perception, decomposition, and inference, but the stage that limits accuracy can change during training. Answer-level rewards neither localize failures to individual stages nor provide discriminative group-relative feedback when all…

301. Long-form Video-to-Text Summarization using Recursive Language Models

     Alessia Intini, Lorenzo Vaiani, Enrico Chen et al. · arXiv:2610.26431

     While VideoLLMs have made significant progress in understanding short video benchmarks, they often struggle to summarize long-form ones. This is mainly due to performance degradation caused by long-context prompts, the inherent difficulty of preserving temporal continuity and a…

302. When Guidance Becomes a Guard: Early Interception of Unsafe Inputs in Text-to-Image Generation

     Yanchun Li, Yujin Wu, Gangkai Shen et al. · arXiv:2610.26452

     Text-to-image diffusion models can produce unsafe content, while output filtering intervenes only after synthesis. We investigate whether native generation signals enable early detection of unsafe and adversarial inputs. We find that classifier-free guidance exposes strong…

303. Progressive multi-scale tumor-state memory for incomplete multimodal brain tumor segmentation

     Milad Rostamian, Mohammadreza Mohammadi, Mohsen Soryani · arXiv:2610.26457

     AI summaryProgressive multi‑scale tumor‑state memory network (PMTM) segments brain tumors from any non‑empty subset of four MRI modalities by iteratively pooling modality‑specific evidence, updating a compact state with learned gating, and guiding residual fusion across scales. It outperforms RobustSeg, RFNet, and mmFormer on BraTS2018, BraTS2020, and BraTS2023 GLI, achieving 75.31%, 77.54%, and 80.47% Dice respectively.

304. YUTELA: A Scalable Edge Cloud AIoT System for Multi-Site Video Anomaly Detection with Scene Specific Adaptation

     Shanle Yao, Babak Rahimi Ardabili, Amir Kimiyaie et al. · arXiv:2610.26468

     Building upon our previous Ancilia framework, this work presents \emph{YUTELA}, an edge--server--cloud AIoT system for long-term, multi-site video anomaly detection. YUTELA replaces workstation-based on-site processing with NVIDIA Jetson Thor edge devices and redesigns the…

305. 3D Diffusion Models for Medical Imaging

     Frank-Peter Schilling, Daniel Barco, Philipp Denzel et al. · arXiv:2610.26475

     Generative deep learning methods such as Variational Autoencoders, Generative Adversarial Networks or Diffusion Models have achieved remarkable results in image synthesis. In particular, Diffusion Models have demonstrated superior image fidelity and sample diversity compared to…

306. PDNet: Exploring Pruning and Distillation for Real-Time Worker Activity Recognition in Industrial Assembly

     Malek Baba, Thinesh Thiyakesan Ponbagavathi, Alina Roitberg · arXiv:2610.26491

     AI summaryPDNet jointly prunes and distills 3D‑CNNs for real‑time assembly HAR. On MECCANO with X3D‑M, pruning alone drops Top‑1 to 25.5%, but PDNet restores it to 39.5% while keeping the model 6.9 MB and 3.78 GFLOPs.

307. Interpretable Face Contrast Modeling for Smartphone Portrait Rendering

     Yujin Cho, Benoît Pochon, Jean-Michel Morel et al. · arXiv:2610.26517

     The quality of smartphone portrait photography has become increasingly important to both manufacturers and users. Recent devices use complex computational pipelines to improve facial rendering under varied lighting. These pipelines vary widely across manufacturers, so the same…

308. Small Average Error, Wide Individual Error: General-Purpose LLMs Estimating Energy Intake from Meal Photos in a Feeding Trial

     Xiaomeng Li · arXiv:2610.26538

     Photo-based estimation of food energy could ease dietary assessment, and general-purpose multimodal large language models (LLMs) make such estimates easy to obtain. Most evaluations use one photo taken before eating and compare with energy served, while dietary assessment needs…

309. Holi-World: Building Holistic World Intelligence from Human-Ego Experience

     Hao Li, Xiyuan Zheng, Yuanyuan Gao et al. · arXiv:2610.26549

     Human egocentric experience offers a rich foundation for learning how the world is organized and how it changes through interaction. Realizing this potential at scale requires turning growing volumes of recordings into learning resources that connect scene structure, object…

310. MRI Reconstruction Risk Estimation from Inference Dynamics

     Tamir Shor, Alex Bronstein · arXiv:2610.26558

     Iterative MRI reconstructors produce intermediate estimates before returning a final image. We study these inference trajectories as a model-internal representation of realized reconstruction risk, measured by log-NMSE. Model-native states are projected to magnitude images,…

311. PrismWAM: Decoupling State Representations in World Action Models

     Tianyu Xin, Wanqi Zhong, Bowen Hu et al. · arXiv:2610.26564

     World Action Models (WAMs) typically represent the world through latent embeddings from pretrained video models, which entangle heterogeneous aspects of the world, from how it looks (e.g., texture and color) to how it is physically configured (e.g., occupancy and pose). This…

312. DenseHap: Dense Hand Haptic Estimation from Monocular Images

     Junhan Xue, Rolandos Alexandros Potamias · arXiv:2610.26583

     As dexterous robotic manipulation advances, scalable annotations of human hand-object interactions are becoming increasingly important. Furthermore, previous hand contact estimation methods suffered from issues such as dorsal-side misclassifications and insufficient precision,…

313. Drive2Gauss: FeedForward 4D Gaussian Reconstruction from Multimodal Generative Driving Videos

     Tong Zhao, Zhiyuan Han, Lin Chen et al. · arXiv:2610.26585

     Generating reusable driving environments requires both controllable future observations and an explicit representation that supports rendering from different viewpoints, yet existing methods still recover explicit scenes from generated RGB videos through optimization for each…

314. Scaling Vision Foundation Models for Toxicologic Pathology

     Jessica Loo, Theodoros Pissas, Byunghak Kang et al. · arXiv:2610.26586

     Here, we present TOX-FM, a family of non-clinical PFMs trained on large-scale diverse histopathology datasets of over one million whole slide images consisting of over 60 organs across six animal species from over 14,000 toxicology studies across six decades. We introduce a…

315. EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience

     Xinhao Ji, Hui En Pang, Hao Li et al. · arXiv:2610.26625

     Reconstructing human activity from egocentric video requires more than recovering individual 3D components: the observer's motion, surrounding geometry, and hand-object interactions must fit together in a consistent world. Achieving this coherence requires both joint…

316. Dual-Scale Condition-Shift Editing for Robust Real-World Image Demoiréing

     Shiyang Zhou, Haijin Zeng, Congpei Qiu et al. · arXiv:2610.26644

     Robust demoiréing of screen-recaptured images remains challenging: moiré varies in frequency, color, and spatial support across display-camera pairs and capture conditions. Dedicated restoration networks trained on finite paired data struggle with this diversity, especially when…

317. SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes?

     Zhaoxi Zhang, Hao Liang, Linxuan Wu et al. · arXiv:2610.26659

     Spatial intelligence requires agents to actively acquire visual evidence and integrate observations into a coherent representation of the three-dimensional world. Yet spatial question answering offers limited evidence of whether models can construct accurate metric maps through…

318. CaptionFuse: Training-Free Composed Product Retrieval Via VLM Intent Captioning and Weighted Embedding Fusion

     Zhao Gao, Mina Metias, Mohamed Noordeen Alaudeen et al. · arXiv:2610.26661

     Composed Image Retrieval (CIR) retrieves products matching a reference image modified by natural language---a core capability for eCommerce search. Existing training-free methods fuse CLIP embeddings via arithmetic, but we identify a fundamental limitation: user text…

319. From Short Clips to Real-Time Streams: Identity-Aware Supervision and Context-Efficient Generation for Video Face Swapping

     Zekai Luo, Yuwei Liang, Ruiqi Ma et al. · arXiv:2610.26681

     Video face swapping transfers the reference identity to a target video while preserving its other attributes and motion. Despite recent progress in reference-guided flow-based methods, jointly achieving faithful identity transfer, high fidelity, and real-time generation over…

320. BrushOff: Stroke-Based Re-rendering as a Watermark Removal Attack on Painting

     Hao Chen, Junnan Xu · arXiv:2610.26682

     Invisible watermarks are used to mark who owns a painting. We show that a simple attack removes the most robust one we test. The attack redraws the watermarked painting with brush strokes, using a pretrained stroke-based renderer. We call this attack BrushOff. It needs no…

321. Foundation Model Driven Optical Flow

     Praroop Chanda, Suryansh Kumar · arXiv:2610.26690

     We investigate how frozen pretrained visual and depth representations support dense optical flow when coupled with a supervised matching network. Building on TTSFlow, this paper studies Dense Prediction Transformer (DPT) feature decoding and one coarse-to-fine refinement stage.…

322. PVBench: Quantifying Physical Plausibility in Video Generation with Object-Level Physical Measurements

     Kaiwei Chen, Moran Li, Zhenye Gan et al. · arXiv:2610.26711

     Recent video generation models produce visually compelling outputs, yet their physical plausibility remains poorly measured. Physical plausibility asks whether generated objects obey geometry, kinematics, dynamics, and conservation laws. Existing benchmarks rely on coarse…

323. A Flexible ROI Analysis Tool for Spatially Consistent Vegetation Index Extraction from Turfgrass Imagery

     Kibon Ku, Thai Thanh Tuan, Jinhyun Ahn · arXiv:2610.26719

     Time-series multispectral imaging is widely used to assess turfgrass condition; however, extracting plot-level vegetation index (VI) values remains labor-intensive when camera positions vary between acquisitions and automated plot tracking is unreliable. We developed an…

324. PlanarCityLoc: Breaking Silhouette Ambiguity in LoD Aerial Localization via 3D Planes

     Oussema Dhaouadi, Shao Jie Hu-Chen, Jakob Gollreiter et al. · arXiv:2610.26729

     Aerial visual localization is crucial for autonomous drone navigation in urban environments where GNSS signals are often degraded. Lightweight textureless Level-of-Detail (LoD) city models provide a geometric map alternative to textured assets. Existing silhouette-based methods…

325. CAVA: Complementary-Label-Aware Data Valuation

     Tan-Ha Mai, Hsuan-Tien Lin, Vu Nguyen · arXiv:2610.26734

     Complementary-label learning (CLL) trains a classifier from labels that name a class an instance does not belong to. Such labels are cheap to collect from crowdworkers or vision-language models (VLMs), making them attractive for large-scale annotation. However, both sources can…

326. CheckDistill: Adaptive Checklist Distillation for Efficient Distributional Reward Modeling in Image Editing

     Yaqi Cai, Zuan Gao, Mengting Chen et al. · arXiv:2610.26740

     Reinforcement learning for instruction-guided image editing requires reward models that provide reliable, accurate, and efficient feedback. Direct scalar prediction compresses multidimensional editing preferences, whereas generative evaluators can expose their judgment through…

327. Beyond Omni: A Unified MLLM for Embodied Sensory Understanding

     Shiyun Xiong, Dongming Wu, Peiwen Sun et al. · arXiv:2610.26751

     Recent embodied systems increasingly build on multimodal large language models (MLLMs) for their pretrained understanding and reasoning capabilities, but their perception of the physical world remains largely limited to basic modalities such as vision and audio. However,…

328. V-JEPA Hand: Simple, Accurate and Fast 3D Hand and Camera Trajectory Recovery from Egocentric Videos

     Ruicheng Li, Qixiu Li, Yu Deng et al. · arXiv:2610.26780

     Egocentric video supports diverse applications, including robot learning, augmented reality, and human activity understanding. Recovering metric 3D hand trajectories from such video requires estimating both hand motion and camera egomotion to express the hands in a consistent…

329. Mise-en-Video: Fine-Grained 3D Control for Multi-Shot Video Generation

     Yingshuang Zou, Xiaoyu Shi, Yawen Luo et al. · arXiv:2610.26801

     Recent advances in text-to-video generation have enabled high-fidelity single-shot synthesis, yet practical storytelling creation requires multi-shot videos with both strong inter-shot consistency and fine-grained controllability. Existing multi-shot video generation approaches…

330. What Language Adds, and at What Cost: Four Controlled Probes of Pathology Vision-Language Models

     Wenhao Zhang, Zhongliang Zhou, Yi Ping et al. · arXiv:2610.26805

     Vision--language models (VLMs) promise to make computational pathology queryable and able to draft reports, but how much language adds beyond the vision backbone is unclear, because almost every system is evaluated by its developers on test sets of their own choosing. We grade…

331. State World Models: Execute the Rules, Generate the Pixels

     Chan Yan Ki, Yanran Zhang, Wenzhao Zheng et al. · arXiv:2610.26811

     Existing world models can generate action-conditioned video from a single image. Many focus on camera movement, allowing users to explore the generated scene. However, navigation alone does not ensure that an action follows persistent rules when its outcome depends on state that…

332. PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models

     Yude Zou, Yifei Yao, Guanbo Jiao et al. · arXiv:2610.26813

     World Action Models (WAMs) jointly predict future video and generate actions, yet remain sensitive to task-irrelevant visual perturbations, i.e., changes in visual appearance that should not affect robot behavior. We identify a gap in conventional WAM training: video prediction…

333. Real-Time Radar Semantic and Panoptic Scene Understanding

     Santosh Parit, Julia Hindel, Rohit Mohan et al. · arXiv:2610.26816

     Robust scene understanding under varying environmental conditions is essential for safe autonomous driving. While cameras and LiDAR degrade in adverse weather and poor lighting, Frequency-Modulated Continuous-Wave (FMCW) spinning radar provides dependable 360{\deg} perception at…

334. Open-Form Geometric Style Transfer

     Adar Elad, Gal Metzer, Omer Dahari et al. · arXiv:2610.26820

     AI summaryThe paper learns a geometric style from stylized Latin glyphs and applies it to unrelated shapes such as Greek characters and human poses. It uses a universal Geometry LoRA trained on diverse synthetic sketches to disentangle shape structure from style, achieving successful cross‑family transfer.

335. RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development

     Yuchi Wang, Rongyao Fang, Weihao Yu et al. · arXiv:2610.26826

     AI summaryRefCraft introduces a benchmark of 300 multimodal tasks evaluating AI agents on webpage development using visual references and text instructions. Experiments show that current frontier models struggle with this workflow, with the top system achieving a score of only 52.12 out of 100.

336. SpecKCast: Spectral Kalman Forecasting for Diffusion Acceleration

     Javad Rajabi, Hakki Can Karaimer, David B. Lindell et al. · arXiv:2610.26837

     Diffusion Transformers (DiTs) achieve high-fidelity image and video generation, but they suffer from substantial inference costs due to repeated transformer evaluations across many denoising steps. Feature caching reduces this cost by reusing features across denoising steps.…

337. Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy

     Zekai Li, Jiayi Yu, Jinlu Zhang et al. · arXiv:2610.26848

     In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and…

338. GenOlympiad: Benchmarking Frontier Image Models at Their Capability Boundaries

     Xi Hu, Martin Bell, Liang Zheng · arXiv:2610.26850

     Text-to-image models have made tremendous progress and can produce high-quality images. However, a question which would be asked in any other discipline is where the frontier models fall short. We introduce GenOlympiad, a benchmark which provides the most challenging text…

339. Manga2Animate: An Agentic Framework for Faithful Manga Animation with Script and Storyboard Planning

     Sheng-Wei Zhang, Hsi-An Chen, Yan-Ting Liou et al. · arXiv:2610.26851

     Adapting existing manga into coherent animation that stays faithful to the source is a valuable yet challenging task. Existing methods either access the manga mainly through text, losing its characters and art style, or feed manga pages directly into video generation models,…

340. InfiniDream: Open-Ended Photorealistic 3D Worlds Through Procedural Generation

     Yi-Chuan Huang, Hao-Jen Chien, Zheng-Hui Huang et al. · arXiv:2610.26853

     Procedural generation plays a key role in synthetic data creation and game development, offering greater controllability, editability, and diversity than static mesh representations. However, VLM-based code generation often produces relatively simple 3D assets, while…

341. Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience

     Tianqi Liu, Hanfei Zhao, Qifeng Liang et al. · arXiv:2610.26855

     World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first…

342. Reinforcement Learning with Visually Attributable On-Policy Self-Distillation

     Shiao Xie, Siyu Chen, Jianwei Lv et al. · arXiv:2610.26856

     Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervision with dense token-level corrections…

343. VideoMaMa++: Temporally Consistent Video Matting via Preserve-and-Refine Token

     Sangbeom Lim, Seoung Wug Oh, Heeji Yoon et al. · arXiv:2610.26857

     Diffusion-based video matting has recently emerged as a leading approach, leveraging video diffusion priors from internet-scale pretraining to achieve strong zero-shot generalization from limited synthetic data. However, the heavy compute of these diffusion backbones limits…

344. OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory

     Yu Zhou, Cheng-Fu Yang, Rui Sun et al. · arXiv:2610.26860

     AI summaryOpenVLHarness adds a tool‑capability layer and persistent multimodal memory, enabling large language models to fetch, interpret, and store heterogeneous evidence across reasoning steps. It boosts performance on 22 benchmarks by 9.7–12.6 points over no‑tool baselines while reducing token cost.

345. Can It Actually Work? Synthesizing Functional CAD Mechanisms

     Wenqian Zhang, Yangyi Huang, Zigeng Xu et al. · arXiv:2610.26861

     While LLM agents can generate visually plausible CAD assemblies, these designs often fail to fulfill their intended physical functions. We study functional CAD design, a setting where agents interpret underspecified requests to construct and iteratively revise mechanisms that…

346. Text-to-Image Generative Modeling Without Language Pretraining

     Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas et al. · arXiv:2610.26863

     AI summaryHalcyon trains multimodal autoregressive models from scratch on 100M image‑text pairs, surpassing state‑of‑the‑art text‑to‑image systems while using far less pretraining compute and showing language pretraining is unnecessary.