This paper introduces a hybrid Mamba-Transformer (MT) framework for remote sensing image super-resolution, aiming to overcome the limitations of CNNs and transformers in capturing long-range dependencies and maintaining computational efficiency. MT combines a focused mamba block (FMB) with a snake vision state-space module (SVSSM) for global feature modeling and a pixel-adaptive block (PAB) for pixel-level multiscale enhancement. Experiments on benchmark datasets demonstrate that MT outperforms state-of-the-art methods, achieving a better trade-off between performance and computational cost, specifically reducing parameters and FLOPs compared to MambaIRv2 while improving PSNR.

The paper introduces the Visual Reasoning Benchmark (VRB), a new dataset of 701 visual reasoning questions sourced from primary school exams in Zambia and India, designed to evaluate multimodal large language models (MLLMs). The VRB focuses on minimal-text images to simulate realistic classroom visual reasoning problems, covering tasks like analogy, pattern completion, and spatial matching. Experiments using the VRB reveal that MLLMs exhibit a "jagged frontier" of capabilities, performing well on static tasks like counting but struggling with dynamic spatial operations like folding and rotation.

This paper introduces CSEval, a framework for evaluating the clinical semantic alignment between text prompts and generated medical images, addressing the limitations of existing metrics focused on realism and diversity. CSEval uses language models to identify semantic inconsistencies related to anatomical location and pathology, demonstrating a correlation with expert clinical judgment. The framework offers a scalable method for assessing the clinical reliability of generated medical images, crucial for the safe deployment of text-to-image models in healthcare.

The paper introduces DeepSight, an open-source toolkit designed to integrate safety evaluation and diagnosis for large language models (LLMs) and multimodal large language models (MLLMs). DeepSight combines DeepSafe, an evaluation toolkit, and DeepScan, a diagnosis toolkit, to provide a more comprehensive safety workflow. By unifying task and data protocols, DeepSight aims to bridge the gap between black-box risk evaluation and white-box mechanistic understanding, facilitating targeted safety alignment.

The paper introduces IncompeBench, a new benchmark for Music Information Retrieval (MIR) consisting of 1,574 permissively licensed music snippets, 500 diverse queries, and over 125,000 relevance judgements. This benchmark addresses the lack of high-quality evaluation datasets in MIR, enabling more rigorous and reproducible research. High inter-annotator agreement was achieved through a multi-stage annotation pipeline, ensuring data quality.

The paper introduces Cross-Modal Robustness Transfer (CMRT) to improve the robustness of End-to-End Speech Translation (E2E-ST) models against morphological variations. CMRT leverages adversarial training in the text modality to transfer robustness to the speech modality, eliminating the need for computationally expensive adversarial speech data generation. Experiments across four language pairs show that CMRT improves adversarial robustness by over 3 BLEU points compared to baseline E2E-ST models.

The paper introduces PosterOmni, a framework for generalized artistic poster creation that tackles both local image editing and global design creation aspects of the task. It achieves this by constructing a multi-task dataset, distilling knowledge from local and global expert models, and applying a unified reward feedback mechanism to align visual fidelity and aesthetic preferences. Experiments on the new PosterOmni-Bench demonstrate that PosterOmni outperforms existing open-source and proprietary systems in reference adherence, composition, and aesthetics.

The paper introduces TexSpot, a diffusion-based texture enhancement framework that addresses view-inconsistency and resolution limitations in 3D texture generation. TexSpot utilizes a novel 3D texture representation called Texlet, which combines point-based and UV-based approaches by encoding local texture patches with a 2D encoder and aggregating them with a 3D encoder. Experiments show that TexSpot significantly improves visual fidelity, geometric consistency, and robustness compared to existing state-of-the-art methods.

The paper introduces RI-Mamba, a rotation-invariant state-space model for text-to-shape retrieval that addresses the limitations of existing methods in handling objects with arbitrary orientations and diverse categories. RI-Mamba disentangles pose from geometry using global and local reference frames and Hilbert sorting to create rotation-invariant token sequences. The model incorporates orientational embeddings via feature-wise linear modulation and employs cross-modal contrastive learning with automated triplet generation for scalable training, achieving state-of-the-art results on the OmniObject3D benchmark.

The paper introduces OMEGA-Avatar, a novel feed-forward framework for generating animatable, 360-degree complete 3D Gaussian head avatars from a single image. To achieve this, the method incorporates a semantic-aware mesh deformation module for improved hair modeling and a multi-view feature splatting module to construct a shared canonical UV representation. Experiments demonstrate that OMEGA-Avatar outperforms existing methods in 360-degree full-head completeness and identity preservation.

The paper investigates modality arbitration in Audio-LLMs, revealing a strong bias towards text over audio when the two modalities conflict, even when audio quality is superior. Using the ALME benchmark, the authors demonstrate that Gemini 2.0 Flash exhibits significantly higher text dominance in audio-text conflicts compared to text-text conflicts. They propose that this text dominance arises from an asymmetry in arbitration accessibility rather than information content, and provide evidence through interventions like forced transcription and fine-tuning ablations.

The paper introduces a supervise-assisted multi-modality fusion diffusion model (MFdiff) to restore standard-dose PET (SPET) images from low-dose PET (LPET) and MR images. MFdiff uses a multi-modality feature fusion module to learn optimized fusion features from MR images and incorporates these features as additional conditions in a diffusion model for iterative SPET image generation. A two-stage supervise-assisted learning strategy leverages both generalized priors from simulated data and specific priors from in-vivo data to improve restoration quality, demonstrating superior performance compared to existing methods.

The paper introduces Spatial Chain-of-Thought (SCoT), a framework that combines the spatial reasoning of Multimodal Large Language Models (MLLMs) with the generative capabilities of diffusion models for improved image generation. SCoT trains a diffusion model on interleaved text-coordinate instructions to enhance layout awareness and uses MLLMs as planners to generate detailed layout plans. Experiments show SCoT achieves state-of-the-art performance on image generation benchmarks and excels in complex reasoning and image editing tasks.

The paper introduces DreamID-Omni, a unified framework for human-centric audio-video generation, addressing tasks like reference-based generation, video editing, and audio-driven animation within a single model. It tackles the challenge of disentangling character identities and voice timbres by employing a Dual-Level Disentanglement strategy and a Symmetric Conditional Diffusion Transformer. Experimental results demonstrate state-of-the-art performance in video, audio, and audio-visual consistency, surpassing even proprietary commercial models.

This paper investigates the use of local vision-language models (VLMs) to improve fine-grained activity recognition in newborn resuscitation videos, comparing them to a TimeSformer baseline. The authors explored zero-shot VLM strategies and fine-tuned VLMs with LoRA on a simulated dataset of 13.26 hours of video. Fine-tuning a local VLM with LoRA achieved an F1 score of 0.91, outperforming the TimeSformer baseline (0.70), suggesting the potential of VLMs for this task.

This paper presents an anatomical analysis of text prompting within vision-language segmentation models, specifically SAM3, revealing significant redundancy in text encoder utilization. Based on these findings, they propose SAM3-LiteText, a lightweight text encoding framework that replaces the original SAM3 text encoder with a compact MobileCLIP student. Experiments demonstrate that SAM3-LiteText reduces text encoder parameters by up to 88% while maintaining segmentation performance on image and video segmentation benchmarks.

This paper introduces ViTaS, a visuomotor learning framework that leverages both visual and tactile information through Soft Fusion Contrastive Learning and a CVAE module to improve performance in manipulation tasks, especially in occluded scenarios. The Soft Fusion Contrastive Learning method is designed to better exploit the alignment and complementarity of visual and tactile representations. Experiments across 12 simulated and 3 real-world environments demonstrate that ViTaS significantly outperforms existing baselines, highlighting the benefits of the proposed fusion and contrastive learning approach.

The paper introduces DeepGen 1.0, a 5B parameter unified multimodal model for image generation and editing, designed to be lightweight and efficient compared to larger models. To enhance semantic understanding in the compact model, they propose Stacked Channel Bridging (SCB) to extract and fuse hierarchical features from VLMs with learnable 'think tokens'. They also employ a three-stage data-centric training strategy, including alignment pre-training, joint supervised fine-tuning, and reinforcement learning with MR-GRPO, achieving state-of-the-art performance on benchmarks like WISE and UniREditBench while using only 50M training samples.

The paper introduces MuRGAt, a new benchmark for evaluating fact-level multimodal attribution in complex reasoning scenarios involving video, audio, and other modalities. MuRGAt requires models to generate answers with explicit reasoning and precise citations that specify modality and temporal segments. The authors also present an automatic evaluation framework that correlates with human judgments, revealing that current MLLMs often hallucinate citations even with correct reasoning, and that increasing reasoning depth can degrade attribution accuracy.

The paper introduces A$^{2}$V-SLP, an alignment-aware variational framework for sign language production that learns disentangled latent distributions for each articulator. This approach uses a disentangled VAE to encode sign pose sequences and extract articulator-specific mean and variance vectors, which then serve as distributional supervision for a non-autoregressive Transformer that predicts latent means and log-variances from text embeddings. By employing stochastic sampling and a gloss attention mechanism, A$^{2}$V-SLP achieves state-of-the-art back-translation performance and enhances motion realism in gloss-free sign language production.

This paper introduces VLAW, an iterative algorithm for co-improving vision-language-action (VLA) policies and action-conditioned video generation world models using real-world rollouts. VLAW leverages real-world data to refine the world model, which is then used to generate synthetic data for further policy improvement, addressing the limitations of world models trained solely on demonstration datasets. Experiments on a real robot demonstrate a 39.2% absolute improvement in success rate over the base policy, highlighting the effectiveness of the iterative co-improvement strategy.

The paper addresses object hallucination in Multimodal Large Language Models (MLLMs) by improving visual contrastive decoding (VCD) through the creation of an object-aligned auxiliary view. This auxiliary view is constructed by masking the most salient visual evidence based on object-centric attention from self-supervised Vision Transformers, thereby disrupting unsupported tokens during decoding. The proposed method, "Mask What Matters," is prompt-agnostic, model-agnostic, and computationally efficient, leading to improved performance on object hallucination benchmarks.

This paper investigates gender and skin-tone biases in Gemini Flash 2.5 Image and GPT Image 1.5 by generating 3,200 images from semantically neutral prompts. Using a pipeline involving color normalization, facial landmark masking, and skin tone quantification via Monk, PERLA, and Fitzpatrick scales, the study reveals a "default white" bias in both models. Gemini favored female-presenting subjects, while GPT favored male-presenting subjects with lighter skin tones, demonstrating that neutral prompts elicit polarized demographic defaults.

This paper explores test-time verification as a method to improve vision-language-action (VLA) alignment, addressing the "intention-action gap" in embodied instruction following. They demonstrate that scaling both rephrased instructions and generated actions at test time enhances sample diversity and improves action selection. The authors introduce CoVer, a contrastive verifier, and a hierarchical verification inference pipeline, showing that this verification approach outperforms scaling policy pre-training on the SIMPLER and PolaRiS benchmarks.

The paper introduces 3DGSNav, a zero-shot object navigation (ZSON) framework that leverages 3D Gaussian Splatting (3DGS) as persistent memory for vision-language models (VLMs) to improve spatial reasoning. 3DGSNav actively constructs a 3DGS representation of the environment and uses trajectory-guided free-viewpoint rendering to generate frontier-aware first-person views, which are then combined with structured visual prompts and Chain-of-Thought prompting to enhance VLM reasoning. Experiments on multiple benchmarks and a quadruped robot show that 3DGSNav achieves competitive performance compared to existing methods.

This paper introduces a French-focused benchmark for PDF-to-Markdown conversion using VLMs, addressing the lack of evaluation datasets for non-English documents and the over-penalization of formatting variations in existing benchmarks. The benchmark consists of challenging French documents selected via model-disagreement sampling and is evaluated using unit-test-style checks targeting specific failure modes like text presence and reading order, combined with category-specific normalization. Results across 15 models show that proprietary models exhibit higher robustness on handwriting and forms, while open-weight models are competitive on standard layouts.

The paper introduces EmoSpace, a framework for emotion-aware content generation that learns dynamic emotion prototypes via vision-language alignment to enable fine-grained emotional control in VR content creation. EmoSpace uses a hierarchical emotion representation with learnable prototypes that evolve during training, allowing for control without explicit emotion labels. Experiments demonstrate EmoSpace's superior performance in emotional image outpainting, stylized generation, and emotional panorama generation, further validated by a user study comparing emotional perception in VR versus desktop environments.

This paper introduces ImagineAgent, a framework that uses cognitive reasoning and generative imagination to improve Open-Vocabulary Human-Object Interaction (OV-HOI) comprehension. ImagineAgent constructs cognitive maps to model relationships between entities and actions, and uses retrieval augmentation, image cropping, and diffusion models to gather knowledge and visual evidence. Experiments on SWIG-HOI and HICO-DET show state-of-the-art performance with significantly less training data.

The paper introduces Hi-SAM, a novel multi-modal recommendation framework designed to address limitations in semantic ID-based approaches, specifically suboptimal tokenization and architecture-data mismatch. Hi-SAM employs a Disentangled Semantic Tokenizer (DST) that uses geometry-aware alignment and coarse-to-fine quantization to separate shared and modality-specific semantics, and a Hierarchical Memory-Anchor Transformer (HMAT) that incorporates hierarchical positional encoding and anchor tokens to better model user-item interactions. Experiments on real-world datasets and a large-scale social platform demonstrate that Hi-SAM outperforms state-of-the-art baselines, particularly in cold-start scenarios, achieving a 6.55% improvement in a core online metric.

The paper introduces Progressive Semantic Illusions, a vector sketching task where a single sketch transforms semantically through sequential stroke additions. They propose Stroke of Surprise, a generative framework using sequence-aware joint optimization with a dual-branch Score Distillation Sampling (SDS) mechanism to satisfy distinct semantic interpretations at different drawing stages. The method dynamically adjusts prefix strokes and uses a novel Overlay Loss to enforce spatial complementarity, achieving superior recognizability and illusion strength compared to baselines.

The paper introduces STVG-R1, a reinforcement learning framework for spatial-temporal video grounding (STVG) that addresses misalignment between textual descriptions and visual coordinates by reformulating per-frame coordinate prediction as instance-level identification using temporally consistent IDs embedded as visual prompts. This approach avoids the need for additional trainable modules and complex alignment strategies. By employing a task-driven reward to optimize temporal accuracy, spatial consistency, and structural format regularization, STVG-R1 achieves state-of-the-art results on multiple STVG benchmarks and demonstrates strong zero-shot generalization capabilities.

The paper introduces AssetFormer, an autoregressive Transformer model for generating modular 3D assets from text descriptions, addressing the need for high-quality, diverse assets in the digital industry. AssetFormer models the generation of 3D assets as a sequence of primitives with constrained design parameters, adapting module sequencing and decoding techniques from language models. Experiments using real-world modular assets demonstrate the model's effectiveness in streamlining asset creation for professional development and UGC scenarios.

This paper introduces a lightweight RGB-D fusion framework to improve the efficiency and accuracy of Segment Anything Models (SAM). They augment EfficientViT-SAM with monocular depth priors generated by a pretrained estimator, fusing depth information mid-level with RGB features using a dedicated depth encoder. Training on only 11.2k samples, the proposed method outperforms EfficientViT-SAM, demonstrating the effectiveness of depth cues as geometric priors for segmentation.

The paper identifies limitations in current Vision-Language-Action (VLA) models stemming from inadequate visual representations learned through language-image contrastive learning or image-based self-supervised learning. It proposes JEPA-VLA, a method that integrates video predictive embeddings (specifically V-JEPA 2) into VLAs to improve environment understanding and policy priors. Experiments on benchmarks like LIBERO and real-robot tasks demonstrate that JEPA-VLA significantly improves performance by leveraging the ability of video predictive embeddings to encode task-relevant temporal dynamics.

The paper introduces Transform Domain Fusion UNet (TD-FusionUNet), a lightweight deep learning model for next-day wildfire spread prediction using multimodal satellite data. The model incorporates trainable Hadamard Transform and Discrete Cosine Transform layers to capture frequency components in orthogonalized latent spaces, along with custom preprocessing techniques for sparse pre-fire masks. Evaluated on the Next-Day Wildfire Spread and WildfireSpreadTS datasets, TD-FusionUNet achieves an F1 score of 0.591 with only 370k parameters, surpassing a ResNet18-based UNet baseline in the WildfireSpreadTS dataset.

The paper introduces SLD-L2S, a novel lip-to-speech (L2S) framework based on a hierarchical subspace latent diffusion model that directly maps visual lip movements to the continuous latent space of a pre-trained neural audio codec, bypassing intermediate representations. The method employs a hierarchical architecture with parallel subspaces and a diffusion convolution block (DiCB) to enhance interactions within and between subspaces. By using reparameterized flow matching, the framework incorporates speech language model (SLM) and semantic losses during training, leading to state-of-the-art generation quality on benchmark datasets.

This paper introduces LLM-DRS, a novel Large Language Model (LLM)-based framework for disaster reconnaissance summarization in structural health monitoring. The framework integrates vision data and metadata from on-site investigations, using deep convolutional neural networks to extract key attributes like damage state and material type. The extracted data, along with carefully designed prompts, are then fed into an LLM to generate summary reports for individual structures or affected regions.

The paper introduces Code2Worlds, a framework for generating 4D dynamic scenes by formulating the task as language-to-simulation code generation. It addresses the challenges of multi-scale context entanglement and the semantic-physical execution gap by using a dual-stream architecture for disentangled object and environment generation, combined with a physics-aware closed-loop mechanism involving a PostProcess Agent and VLM-Motion Critic. Experiments on the Code4D benchmark demonstrate that Code2Worlds significantly outperforms existing methods in scene generation score (SGS) and richness, while also generating more physically plausible dynamics.

The paper introduces HoloBrain-0, a Vision-Language-Action (VLA) framework designed to improve real-world robot deployment by incorporating robot embodiment priors like multi-view camera parameters and URDF into its architecture. They employ a "pre-train then post-train" paradigm, achieving SOTA results on simulation benchmarks and strong performance on real-world manipulation tasks, even with a small 0.2B-parameter variant. The authors open-source the entire HoloBrain ecosystem, including pre-trained models, post-trained checkpoints, and a full-stack VLA infrastructure called RoboOrchard, to facilitate research and adoption.

This paper introduces ReNoV, a diffusion-based novel view synthesis framework that leverages external visual representations as conditions to improve geometric consistency. The authors analyze the correspondence capabilities within spatial attention of external representations and use this to guide the diffusion process via representation projection modules. Experiments demonstrate that ReNoV achieves significant improvements in reconstruction fidelity and inpainting quality compared to existing diffusion-based methods, particularly in scenarios with sparse, unposed images.

The paper introduces GigaBrain-0.5M*, a vision-language-action (VLA) model trained using world model-based reinforcement learning to improve multi-step action prediction. They leverage the spatiotemporal reasoning capabilities of video world models pre-trained on large video datasets to enhance VLA learning. By integrating world model-based reinforcement learning via RAMP (Reinforcement leArning via world Model-conditioned Policy), GigaBrain-0.5M* achieves significant performance gains (approximately 30%) over the RECAP baseline on complex manipulation tasks and demonstrates reliable long-horizon execution in real-world deployments.

This paper introduces a method for learning structured latent representations in RL where distances reflect transition costs, providing a geometric interpretation of uncertainty without explicit probabilistic modeling. They achieve this with a multimodal latent transition model and inverse distance weighting for sensor fusion, enabling adaptive integration of multiple sensor modalities. Empirical validation on multimodal RL tasks demonstrates improved robustness to sensor noise, superior state estimation, and enhanced RL agent performance compared to baselines, eliminating the need for noise augmentation.

The paper introduces MTL-VQA, a multi-task learning framework for no-reference video quality assessment (NR-VQA) of gaming videos, addressing the scarcity of human-rated data by leveraging full-reference (FR) metrics as supervisory signals. By adaptively weighting and jointly optimizing multiple FR objectives during pretraining, the method learns shared perceptual representations relevant to video quality. Experiments demonstrate that MTL-VQA achieves competitive performance compared to state-of-the-art NR-VQA methods in both MOS-supervised and label-efficient settings.

The paper addresses the problem of biased uncertainty estimation in Test-Time Adaptation (TTA) of vision-language models like CLIP, which arises from pre-training on imbalanced web data. They propose Adaptive Debiasing Tsallis Entropy (ADTE), a generalization of Shannon Entropy that incorporates a class-specific parameter to account for label bias estimated from incoming test instances. ADTE outperforms state-of-the-art TTA methods on ImageNet variants and cross-domain benchmarks by accurately selecting high-confidence views and integrating with a label adjustment strategy.

The paper introduces UniT, a framework for multimodal chain-of-thought test-time scaling that allows a unified model to iteratively reason, verify, and refine its outputs. UniT employs agentic data synthesis to create training data, trains a unified model, and uses flexible test-time inference to encourage cognitive behaviors. Experiments demonstrate that models trained on short reasoning trajectories generalize to longer inference chains, sequential chain-of-thought reasoning is more scalable than parallel sampling, and training on generation/editing trajectories improves out-of-distribution visual reasoning.

The paper introduces Robot-DIFT, a framework that distills geometric priors from a frozen diffusion model into a deterministic Spatial-Semantic Feature Pyramid Network (S2-FPN) to improve visuomotor control. This distillation process aims to address the structural mismatch between vision encoders optimized for semantic invariance and the geometric sensitivity required for precise manipulation. Robot-DIFT, pretrained on the DROID dataset, achieves superior geometric consistency and control performance compared to discriminative baselines by leveraging the geometric dependencies encoded within diffusion model latent manifolds.

The paper introduces ABot-N0, a Vision-Language-Action (VLA) foundation model designed for unified embodied navigation across five core tasks: Point-Goal, Object-Goal, Instruction-Following, POI-Goal, and Person-Following. ABot-N0 employs a hierarchical "Brain-Action" architecture, combining an LLM-based cognitive brain for semantic reasoning with a Flow Matching-based action expert for trajectory generation. The model is trained on a large-scale dataset of 16.9M expert trajectories and 5.0M reasoning samples, achieving state-of-the-art performance on seven benchmarks and demonstrating robust long-horizon navigation in real-world environments.

The paper introduces WebTestPilot, an LLM-based agent for end-to-end web testing against natural language specifications that addresses the challenges of implicit oracle inference and probabilistic reasoning. WebTestPilot uses a symbolization layer to represent GUI elements as symbols and translates natural language into step-by-step instructions with inferred pre- and post-conditions over these symbols, effectively capturing data, temporal, and causal dependencies for validation. Experiments on a new benchmark of bug-injected web applications demonstrate that WebTestPilot achieves a 99% task completion rate with 96% precision and 96% recall in bug detection, significantly outperforming existing LLM-based approaches.

The paper introduces DiffPlace, a diffusion-based framework for generating place-controllable street view images from text, BEV maps, and object bounding boxes, specifically addressing the challenge of generating background-consistent urban scenes. DiffPlace employs a place-ID controller, using linear projection, a perceiver transformer, and contrastive learning to map place-ID embeddings into a CLIP space, enabling control over background consistency while allowing foreground variations. Experiments demonstrate that DiffPlace achieves superior generation quality and improves visual place recognition performance when used for data augmentation compared to existing methods.

The paper introduces ScalSelect, a training-free multimodal data selection method for visual instruction tuning (VIT) that addresses the computational expense and redundancy of large-scale datasets. ScalSelect constructs sample representations by extracting visual features most attended by instruction tokens in the target VLM and then identifies samples whose representations best approximate the dominant subspace of the full dataset. Experiments demonstrate that ScalSelect achieves comparable or superior performance to full-data training using significantly less data (e.g., 16%).