Vision-language models trained on large-scale multimodal datasets show strong demographic biases, but the role of training data in producing these biases remains unclear. A major barrier has been the lack of demographic annotations in web-scale datasets such as LAION-400M. We address this gap by creating person-centric annotations for the full dataset, including over 276 million bounding boxes, perceived gender and race/ethnicity labels, and automatically generated captions. These annotations are produced through validated automatic labeling pipelines combining object detection, multimodal captioning, and finetuned classifiers. Using them, we uncover demographic imbalances and harmful associations, such as the disproportionate linking of men and individuals perceived as Black or Middle Eastern with crime-related and negative content. We also show that a linear fit predicts 60-70% of gender bias in CLIP and Stable Diffusion from direct co-occurrences in the data. Our resources establish the first large-scale empirical link between dataset composition and downstream model bias. Code is available at https://github.com/ExplainableML/LAION-400M-Person-Centric-Annotations.

Reasoning has emerged as a pivotal capability in Large Language Models (LLMs). Through Reinforcement Learning (RL), typically Group Relative Policy Optimization (GRPO), these models are able to solve complex tasks such as mathematics and code generation. Building on these advances, recent research has sought to extend reasoning to Vision-Language Models (VLMs), yielding promising results across diverse visual tasks. Despite this progress, our study uncovers the dual nature of multimodal reasoning: while it substantially enhances logical inference and facilitates performance on challenging problems, it may gradually impair perceptual grounding, leading to recognition failures on otherwise basic visual questions. Through further analysis, we attribute this phenomenon to visual forgetting, wherein prolonged reasoning causes the model to increasingly disregard visual input. To address this, we propose Vision-Anchored Policy Optimization (VAPO), a simple yet effective method that explicitly steers the reasoning process toward visually grounded trajectories. Our result model, VAPO-Thinker-7B, significantly strengthens the model's reliance on visual information and achieves new state-of-the-art results on a wide range of established benchmarks. Project page: https://xytian1008.github.io/VAPO/

We study distributed reinforcement learning (RL) with policy gradient methods under asynchronous and parallel computations and communications. While non-distributed methods are well understood theoretically and have achieved remarkable empirical success, their distributed counterparts remain less explored, particularly in the presence of heterogeneous asynchronous computations and communication bottlenecks. We introduce two new algorithms, Rennala NIGT and Malenia NIGT, which implement asynchronous policy gradient aggregation and achieve state-of-the-art efficiency. In the homogeneous setting, Rennala NIGT provably improves the total computational and communication complexity while supporting the AllReduce operation. In the heterogeneous setting, Malenia NIGT simultaneously handles asynchronous computations and heterogeneous environments with strictly better theoretical guarantees. Our results are further corroborated by experiments, showing that our methods significantly outperform prior approaches.

As large language models evolve, Machine Unlearning has emerged to address growing concerns around user privacy, copyright infringement, and overall safety. Yet state-of-the-art (SOTA) unlearning methods often suffer from catastrophic forgetting and metric imbalance, for example, by over-optimizing one objective (e.g., unlearning effectiveness, utility preservation, or privacy protection) at the expense of others. In addition, small perturbations in the representation or parameter space can be exploited by relearn and jailbreak attacks. To address these challenges, we propose PRISM, a unified framework that enforces dual-space smoothness in representation and parameter spaces to improve robustness and balance unlearning metrics. PRISM consists of two smoothness optimization stages: (i) a representation space stage that employs a robustly trained probe to defend against jailbreak attacks, and (ii) a parameter-space stage that decouples retain-forget gradient conflicts, reduces imbalance, and smooths the parameter space to mitigate relearning attacks. Extensive experiments on WMDP and MUSE, across conversational-dialogue and continuous-text settings, show that PRISM outperforms SOTA baselines under multiple attacks while achieving a better balance among key metrics.

Effective credit risk management is fundamental to financial decision-making, requiring robust models to predict default probabilities and classify financial entities. Traditional machine learning approaches face significant challenges when confronted with high-dimensional data, limited interpretability, rare-event detection, and multi-class risk imbalance. This research proposes a comprehensive multi-stage ensemble pipeline that synthesizes multiple complementary models: econometric models including Ordered logit and ordered probit, supervised learning algorithms, including XGBoost, Random Forest, Support Vector Machine, and Decision Tree; unsupervised methods such as K-Nearest Neighbors; deep learning architectures like Multilayer Perceptron; alongside LASSO regularization for feature selection and dimensionality reduction; and Error-Correcting Output Codes as an Ensemble classifier for handling imbalanced multi-class problems. We implement Permutation Feature Importance analysis for each prediction class across all constituent models to enhance model transparency. Our framework can optimize predictive performance while providing a more holistic approach to credit risk assessment. This research contributes to the development of more accurate and reliable computational models for strategic financial decision support by addressing three fundamental challenges in credit risk modeling. The empirical validation of our approach involves analyzing the Corporate Credit Ratings dataset, which contains credit ratings for 2,029 publicly listed US companies. Results demonstrate that our multi-stage ensemble pipeline significantly enhances the accuracy of financial entity classification regarding credit rating migrations (upgrades and downgrades) and default probability estimation.

A primary bottleneck in large-scale text-to-video generation today is physical consistency and controllability. Despite recent advances, state-of-the-art models often produce unrealistic motions, such as objects falling upward, or abrupt changes in velocity and direction. Moreover, these models lack precise parameter control, struggling to generate physically consistent dynamics under different initial conditions. We argue that this fundamental limitation stems from current models learning motion distributions solely from appearance, while lacking an understanding of the underlying dynamics. In this work, we propose NewtonGen, a framework that integrates data-driven synthesis with learnable physical principles. At its core lies trainable Neural Newtonian Dynamics (NND), which can model and predict a variety of Newtonian motions, thereby injecting latent dynamical constraints into the video generation process. By jointly leveraging data priors and dynamical guidance, NewtonGen enables physically consistent video synthesis with precise parameter control. All data and code are available at https://github.com/pandayuanyu/NewtonGen

Motion blur reduces the clarity of fast-moving objects, posing challenges for detection systems, especially in racket sports, where balls often appear as streaks rather than distinct points. Existing labeling conventions mark the ball at the leading edge of the blur, introducing asymmetry and ignoring valuable motion cues correlated with velocity. This paper introduces a new labeling strategy that places the ball at the center of the blur streak and explicitly annotates blur attributes. Using this convention, we release a new table tennis ball detection dataset. We demonstrate that this labeling approach consistently enhances detection performance across various models. Furthermore, we introduce BlurBall, a model that jointly estimates ball position and motion blur attributes. By incorporating attention mechanisms such as Squeeze-and-Excitation over multi-frame inputs, we achieve state-of-the-art results in ball detection. Leveraging blur not only improves detection accuracy but also enables more reliable trajectory prediction, benefiting real-time sports analytics.

Many practical applications of multi-objective optimization (MOO), including engineering design, autonomous systems, and machine learning, often yield complex Pareto frontiers (e.g., discontinuous, degenerate, or non-convex), which pose challenges for traditional scalarization and Pareto Set Learning (PSL) methods that struggle to approximate them accurately. In this paper, we propose GaussianPSL, a novel framework that uses soft partitions of the Pareto decision/objective space to address the challenges posed by complex Pareto frontiers. Our method dynamically partitions the space, enabling simple MLP networks to learn localized features within each region and then aggregate this information for the final prediction. This partition-aware strategy enhances both exploration and convergence, reduces sensitivity to initialization, and improves robustness against local optima. Experimental results demonstrate that the proposed approach consistently outperforms standard PSL models in learning complex Pareto fronts while maintaining model simplicity. Overall, GaussianPSL offers a new direction for effective, scalable MOO in challenging frontier geometries.

Wide field-of-view (FoV) LiDAR sensors provide dense geometry across large environments, but existing LiDAR-inertial-visual odometry (LIVO) systems generally rely on a single camera, limiting their ability to fully exploit LiDAR-derived depth for photometric alignment and scene colorization. We present Omni-LIVO, a tightly coupled multi-camera LIVO system that leverages multi-view observations to comprehensively utilize LiDAR geometric information across extended spatial regions. Omni-LIVO introduces a Cross-View direct alignment strategy that maintains photometric consistency across non-overlapping views, and extends the Error-State Iterated Kalman Filter (ESIKF) with multi-view updates and adaptive covariance. The system is evaluated on public benchmarks and our custom dataset, showing improved accuracy and robustness over state-of-the-art LIVO, LIO, and visual-inertial SLAM baselines. Code and dataset will be released upon publication.

Diffusion models excel at generating high-quality, diverse images but suffer from training data memorization, raising critical privacy and safety concerns. Data unlearning has emerged to mitigate this issue by removing the influence of specific data without retraining from scratch. We propose ReTrack, a fast and effective data unlearning method for diffusion models. ReTrack employs importance sampling to construct a more efficient fine-tuning loss, which we approximate by retaining only dominant terms. This yields an interpretable objective that redirects denoising trajectories toward the $k$-nearest neighbors, enabling efficient unlearning while preserving generative quality. Experiments on MNIST T-Shirt, CelebA-HQ, CIFAR-10, and Stable Diffusion show that ReTrack achieves state-of-the-art performance, striking the best trade-off between unlearning strength and generation quality preservation.

AI systems often struggle to provide reliable predictions across all inputs, motivating hybrid human-AI decision-making. Existing Learning to Defer (L2D) approaches address this by training models to selectively defer to human experts. However, these approaches require extensive training data annotated by all experts and are sensitive to changes in expert composition, necessitating costly retraining. We propose a training-free, model- and expert-agnostic framework for expert deferral based on conformal prediction. Our method leverages prediction sets from a conformal predictor to quantify label-specific uncertainty and selects the most suitable expert using a segregativity criterion, which measures how well an expert discriminates among plausible labels. Experiments across three models on CIFAR10-H and HAM10000 demonstrate that our method can reduce the number of training labels per expert by up to 91.3% while maintaining predictive accuracy in low-data regimes. Being training-free, it also reduces training time by two orders of magnitude, offering a scalable, alternative to L2D for real-world human-AI collaboration.

Electronic Dance Music (EDM) classification typically relies on industry-defined taxonomies, with current supervised approaches naturally assuming the validity of prescribed subgenre labels. However, whether these commercial distinctions reflect genuine acoustic differences remains largely unexplored. In this paper, we propose an unsupervised approach to discover the natural acoustic structure of EDM independent of commercial labels. To address the historical lack of EDM-specific feature design in MIR, we systematically construct a tailored, interpretable acoustic feature space capturing the genre's defining production techniques, spectral textures, and layered rhythmic patterns. To ensure our findings reflect inherent acoustic structure rather than feature engineering artifacts, we validate our clustering against state-of-the-art pre-trained audio embeddings (MERT and CLAP). Across both our bespoke feature space and the pre-trained embeddings, clustering consistently identifies 20 or fewer natural acoustic families -- suggesting current commercial EDM taxonomy is acoustically overspecified by nearly one-half.

Recently, learned image compression has attracted considerable attention due to its superior performance over traditional methods. However, most existing approaches employ a single entropy model to estimate the probability distribution of pixel values across the entire image, which limits their ability to capture the diverse statistical characteristics of different semantic regions. To overcome this limitation, we propose Segmentation-Assisted Multi-Entropy Models for Lossless Image Compression (SEEC). Our framework utilizes semantic segmentation to guide the selection and adaptation of multiple entropy models, enabling more accurate probability distribution estimation for distinct semantic regions. Experimental results on benchmark datasets demonstrate that SEEC achieves state-of-the-art compression ratios while introducing only minimal encoding and decoding latency. With superior performance, the proposed model also supports Regions of Interest (ROIs) coding condition on the provided segmentation mask. Our code is available at https://github.com/chunbaobao/SEEC.

This paper introduces a scalable paradigm for supervised style transfer by inverting the problem: instead of learning to stylize directly, we learn to destylize, reducing stylistic elements from artistic images to recover their natural counterparts and thereby producing authentic, pixel-aligned training pairs at scale. To realize this paradigm, we propose DeStylePipe, a progressive, multi-stage destylization framework that begins with global general destylization, advances to category-wise instruction adaptation, and ultimately deploys specialized model adaptation for complex styles that prompt engineering alone cannot handle. Tightly integrated into this pipeline, DestyleCoT-Filter employs Chain-of-Thought reasoning to assess content preservation and style removal at each stage, routing challenging samples forward while discarding persistently low-quality pairs. Built on this framework, we construct DeStyle-350K, a large-scale dataset aligning diverse artistic styles with their underlying content. We further introduce BCS-Bench, a benchmark featuring balanced content generality and style diversity for systematic evaluation. Extensive experiments demonstrate that models trained on DeStyle-350K achieve superior stylization quality, validating destylization as a reliable and scalable supervision paradigm for style transfer.

Language-vision understanding has driven the development of advanced perception systems, most notably the emerging paradigm of Referring Multi-Object Tracking (RMOT). By leveraging natural-language queries, RMOT systems can selectively track objects that satisfy a given semantic description, guided through Transformer-based spatial-temporal reasoning modules. End-to-End (E2E) RMOT models further unify feature extraction, temporal memory, and spatial reasoning within a Transformer backbone, enabling long-range spatial-temporal modeling over fused textual-visual representations. Despite these advances, the reliability and robustness of RMOT remain underexplored. In this paper, we examine the security implications of RMOT systems from a design-logic perspective, identifying adversarial vulnerabilities that compromise both the linguistic-visual referring and track-object matching components. Additionally, we uncover a novel vulnerability in advanced RMOT models employing FIFO-based memory, whereby targeted and consistent attacks on their spatial-temporal reasoning introduce errors that persist within the history buffer over multiple subsequent frames. We present VEIL, a novel adversarial framework designed to disrupt the unified referring-matching mechanisms of RMOT models. We show that carefully crafted digital and physical perturbations can corrupt the tracking logic reliability, inducing track ID switches and terminations. We conduct comprehensive evaluations using the Refer-KITTI dataset to validate the effectiveness of VEIL and demonstrate the urgent need for security-aware RMOT designs for critical large-scale applications.

Despite advances in the Transformer architecture, their effectiveness for long-term time series forecasting (LTSF) remains controversial. In this paper, we investigate the potential of integrating explicit periodicity modeling into the self-attention mechanism to enhance the performance of Transformer-based architectures for LTSF. Specifically, we propose PENGUIN, a simple yet effective periodic-nested group attention mechanism. Our approach introduces a periodic-aware relative attention bias to directly capture periodic structures and a grouped multi-query attention mechanism to handle multiple coexisting periodicities (e.g., daily and weekly cycles) within time series data. Extensive experiments across diverse benchmarks demonstrate that PENGUIN consistently outperforms both MLP-based and Transformer-based models. Code is available at https://github.com/ysygMhdxw/AISTATS2026_PENGUIN.

People control their bodies to establish contact with the environment. To comprehensively understand actions across diverse visual contexts, it is essential to simultaneously consider \textbf{what} action is occurring and \textbf{where} it is happening. Current methodologies, however, often inadequately capture this duality, typically failing to jointly model both action semantics and their spatial contextualization within scenes. To bridge this gap, we introduce a novel vision task that simultaneously predicts high-level action semantics and fine-grained body-part contact regions. Our proposed framework, PaIR-Net, comprises three key components: the Contact Prior Aware Module (CPAM) for identifying contact-relevant body parts, the Prior-Guided Concat Segmenter (PGCS) for pixel-wise contact segmentation, and the Interaction Inference Module (IIM) responsible for integrating global interaction relationships. To facilitate this task, we present PaIR (Part-aware Interaction Representation), a comprehensive dataset containing 13,979 images that encompass 654 actions, 80 object categories, and 17 body parts. Experimental evaluation demonstrates that PaIR-Net significantly outperforms baseline approaches, while ablation studies confirm the efficacy of each architectural component. The code and dataset will be released upon publication.

Supervised fine-tuning (SFT) plays a critical role for pretrained large language models (LLMs), notably enhancing their capacity to acquire domain-specific knowledge while preserving or potentially augmenting their general-purpose capabilities. However, the efficacy of SFT hinges on data quality as well as data volume, otherwise it may result in limited performance gains or even degradation relative to the associated baselines. To mitigate such reliance, we suggest categorizing tokens within each corpus into two parts -- positive and negative tokens -- based on whether they are useful to improve model performance. Positive tokens can be trained in common ways, whereas negative tokens, which may lack essential semantics or be misleading, should be explicitly forgotten. Overall, the token categorization facilitates the model to learn less informative messages, and the forgetting guides the model on what information to learn more precisely. We conduct experiments across diverse and well-established benchmarks using various model architectures, demonstrating that this forgetting mechanism enhances model performance.

Multimodal reasoning over long-horizon video is challenging due to the need for precise spatiotemporal fusion and alignment across modalities. While recent methods such as Group Relative Policy Optimization (GRPO) have shown promise in this domain, they suffer from three key limitations: (1) data inefficiency from their on-policy design, (2) a vanishing advantage problem, where identical or near-identical rewards within a group eliminate the learning signal by producing zero-valued advantages, and (3) uniform credit assignment that fails to emphasize critical reasoning steps. We introduce $\textbf{AVATAR}$ ($\textbf{A}$udio-$\textbf{V}$ideo $\textbf{A}$gen$\textbf{t}$ for $\textbf{A}$lignment and $\textbf{R}$easoning), a framework that addresses these limitations through two core components: (1) an off-policy training architecture that improves sample efficiency and resolves vanishing advantages by reusing past experiences with greater reward diversity, and (2) Temporal Advantage Shaping (TAS), a credit assignment strategy that emphasizes early (planning) and late (synthesis) reasoning phases. $\textbf{AVATAR}$ achieves strong performance across various benchmarks, outperforming the Qwen2.5-Omni baseline by $\mathbf{+5.4}$ on MMVU, $\mathbf{+4.9}$ on OmniBench, and $\mathbf{+4.5}$ on Video-Holmes. Furthermore, it surpasses standard GRPO by $\mathbf{+3.7}$ on OmniBench and $\mathbf{+1.9}$ on Video-Holmes, while demonstrating $\textbf{$5$$\times$ sample efficiency}$, requiring $80\%$ fewer generated completions to reach target performance.

Automated bioacoustic analysis is essential for biodiversity monitoring and conservation, requiring advanced deep learning models that can adapt to diverse bioacoustic tasks. This article presents a comprehensive review of large-scale pretrained bioacoustic foundation models and systematically investigates their transferability across multiple bioacoustic classification tasks. We overview bioacoustic representation learning by analysing pretraining data sources and benchmarks. On this basis, we review bioacoustic foundation models, dissecting the models' training data, preprocessing, augmentations, architecture, and training paradigm. Additionally, we conduct an extensive empirical study of selected models on the BEANS and BirdSet benchmarks, evaluating generalisability under linear and attentive probing. Our experimental analysis reveals that Perch~2.0 achieves the highest BirdSet score (restricted evaluation) and the strongest linear probing result on BEANS, building on diverse multi-taxa supervised pretraining; that BirdMAE is the best model among probing-based strategies on BirdSet and second on BEANS after BEATs$_{NLM}$, the encoder of NatureLM-audio; that attentive probing is beneficial to extract the full performance of transformer-based models; and that general-purpose audio models trained with self-supervised learning on AudioSet outperform many specialised bird sound models on BEANS when evaluated with attentive probing. These findings provide valuable guidance for practitioners selecting appropriate models to adapt them to new bioacoustic classification tasks via probing.

Robotic middleware is fundamental to ensuring reliable communication among system components and is crucial for intelligent robotics, autonomous vehicles, and smart manufacturing. However, existing robotic middleware often struggles to meet the diverse communication demands, optimize data transmission efficiency, and maintain scheduling determinism between Orin computing units in large-scale L4 autonomous vehicle deployments. This paper presents RIMAOS2C, a service discovery-based hybrid network communication middleware designed to tackle these challenges. By leveraging multi-level service discovery multicast, RIMAOS2C supports a wide variety of communication modes, including multiple cross-chip Ethernet protocols and PCIe communication capabilities. Its core mechanism, the Message Bridge, optimizes data flow forwarding and employs shared memory for centralized message distribution, reducing message redundancy and minimizing transmission delay uncertainty. Tested on L4 vehicles and Jetson Orin domain controllers, RIMAOS2C leverages TCP-based ZeroMQ to overcome the large-message transmission bottleneck in native CyberRT. In scenarios with two cross-chip subscribers, it eliminates message redundancy and improves large-data transmission efficiency by 36 to 40 percent while reducing callback latency variation by 42 to 906 percent. This research advances the communication capabilities of robotic operating systems and proposes a novel approach to optimizing communication in distributed computing architectures for autonomous driving.

As large reasoning models (LRMs) grow more capable, chain-of-thought (CoT) reasoning introduces new safety challenges. Existing SFT-based safety alignment studies dominantly focused on filtering prompts with safe, high-quality responses, while overlooking hard prompts that always elicit harmful outputs. To fill this gap, we introduce UnsafeChain, a safety alignment dataset constructed from hard prompts with diverse sources, where unsafe completions are identified and explicitly corrected into safe responses. By exposing models to unsafe behaviors and guiding their correction, UnsafeChain enhances safety while preserving general reasoning ability. We fine-tune three LRMs on UnsafeChain and compare them against recent SafeChain and STAR-1 across six out-of-distribution and five in-distribution benchmarks. UnsafeChain consistently outperforms prior datasets, with even a 1K subset matching or surpassing baseline performance, demonstrating the effectiveness and generalizability of correction-based supervision. We release our dataset and code at https://github.com/mbzuai-nlp/UnsafeChain

Deep learning typically relies on end-to-end backpropagation for training, a method that inherently suffers from issues such as update locking during parameter optimization, high GPU memory consumption, and a lack of biological plausibility. In contrast, supervised local learning seeks to mitigate these challenges by partitioning the network into multiple local blocks and designing independent auxiliary networks to update each block separately. However, because gradients are propagated solely within individual local blocks, performance degradation occurs, preventing supervised local learning from supplanting end-to-end backpropagation. To address these limitations and facilitate inter-block information flow, we propose the Momentum Auxiliary Network++ (MAN++). MAN++ introduces a dynamic interaction mechanism by employing the Exponential Moving Average (EMA) of parameters from adjacent blocks to enhance communication across the network. The auxiliary network, updated via EMA, effectively bridges the information gap between blocks. Notably, we observed that directly applying EMA parameters can be suboptimal due to feature discrepancies between local blocks. To resolve this issue, we introduce a learnable scaling bias that balances feature differences, thereby further improving performance. We validate MAN++ through extensive experiments on tasks that include image classification, object detection, and image segmentation, utilizing multiple network architectures. The experimental results demonstrate that MAN++ achieves performance comparable to end-to-end training while significantly reducing GPU memory usage. Consequently, MAN++ offers a novel perspective for supervised local learning and presents a viable alternative to conventional training methods.

We present a novel approach to compute three-dimensional Magnetohydrodynamic equilibria by parametrizing Fourier modes with artificial neural networks and compare it to equilibria computed by conventional solvers. The full nonlinear global force residual across the volume in real space is then minimized with first order optimizers. Already,we observe competitive computational cost to arrive at the same minimum residuals computed by existing codes. With increased computational cost,lower minima of the residual are achieved by the neural networks,establishing a new lower bound for the force residual. We use minimally complex neural networks,and we expect significant improvements for solving not only single equilibria with neural networks,but also for computing neural network models valid over continuous distributions of equilibria.

Cinematography, the fundamental visual language of film, is essential for conveying narrative, emotion, and aesthetic quality. While recent Vision-Language Models (VLMs) demonstrate strong general visual understanding, their proficiency in comprehending the nuanced cinematic grammar embedded within individual shots remains largely unexplored and lacks robust evaluation. This critical gap limits both fine-grained visual comprehension and the precision of AI-assisted video generation. To address this, we introduce ShotBench, a comprehensive benchmark specifically designed for cinematic language understanding. It features over 3.5k expert-annotated QA pairs from images and video clips, meticulously curated from over 200 acclaimed (predominantly Oscar-nominated) films and spanning eight key cinematography dimensions. Our evaluation of 24 leading VLMs on ShotBench reveals their substantial limitations: even the top-performing model achieves less than 60% average accuracy, particularly struggling with fine-grained visual cues and complex spatial reasoning. To catalyze advancement in this domain, we construct ShotQA, a large-scale multimodal dataset comprising approximately 70k cinematic QA pairs. Leveraging ShotQA, we develop ShotVL through supervised fine-tuning and Group Relative Policy Optimization. ShotVL significantly outperforms all existing open-source and proprietary models on ShotBench, establishing new state-of-the-art performance. We open-source our models, data, and code to foster rapid progress in this crucial area of AI-driven cinematic understanding and generation.

Despite significant advancements in Vision-Language Models (VLMs), the performance of existing VLMs remains hindered by object hallucination, a critical challenge to achieving accurate visual understanding. To address this issue, we propose SECOND: Selective and Contrastive Decoding, a novel approach that enables VLMs to effectively leverage multi-scale visual information with an object-centric manner, closely aligning with human visual perception. SECOND progressively selects and integrates multi-scale visual information, facilitating a more precise interpretation of images. By contrasting these visual information iteratively, SECOND significantly reduces perceptual hallucinations and outperforms a wide range of benchmarks. Our theoretical analysis and experiments highlight the largely unexplored potential of multi-scale application in VLMs, showing that prioritizing and contrasting across scales outperforms existing methods.

Patients have distinct information needs about their hospitalization that can be addressed using clinical evidence from electronic health records (EHRs). While artificial intelligence (AI) systems show promise in meeting these needs, robust datasets are needed to evaluate the factual accuracy and relevance of AI-generated responses. To our knowledge, no existing dataset captures patient information needs in the context of their EHRs. We introduce ArchEHR-QA, an expert-annotated dataset based on real-world patient cases from intensive care unit and emergency department settings. The cases comprise questions posed by patients to public health forums, clinician-interpreted counterparts, relevant clinical note excerpts with sentence-level relevance annotations, and clinician-authored answers. To establish benchmarks for grounded EHR question answering (QA), we evaluated three open-weight large language models (LLMs)--Llama 4, Llama 3, and Mixtral--across three prompting strategies: generating (1) answers with citations to clinical note sentences, (2) answers before citations, and (3) answers from filtered citations. We assessed performance on two dimensions: Factuality (overlap between cited note sentences and ground truth) and Relevance (textual and semantic similarity between system and reference answers). The final dataset contains 134 patient cases. The answer-first prompting approach consistently performed best, with Llama 4 achieving the highest scores. Manual error analysis supported these findings and revealed common issues such as omitted key clinical evidence and contradictory or hallucinated content. Overall, ArchEHR-QA provides a strong benchmark for developing and evaluating patient-centered EHR QA systems, underscoring the need for further progress toward generating factual and relevant responses in clinical contexts.

Story visualization aims to generate coherent image sequences that faithfully represent a narrative and match given character references. Despite progress in generative models, existing benchmarks remain narrow in scope, often limited to short prompts, lacking character references, or single-image cases, failing to reflect real-world narrative complexity and obscuring true model performance.We introduce ViStoryBench, a comprehensive benchmark designed to evaluate story visualization models across varied narrative structures, visual styles, and character settings. It features richly annotated multi-shot scripts derived from curated stories spanning literature, film, and folklore. Large language models assist in story summarization and script generation, with all outputs verified by humans for coherence and fidelity. Character references are carefully curated to maintain consistency across different artistic styles. ViStoryBench proposes a suite of multi-dimensional automated metrics to evaluate character consistency, style similarity, prompt alignment, aesthetic quality, and artifacts like copy-paste behavior. These metrics are validated through human studies and used to assess a broad range of open-source and commercial models, enabling systematic analysis and encouraging advances in visual storytelling.

Latent Video Diffusion Models (LVDMs) have achieved state-of-the-art generative quality for image and video generation; however, they remain brittle under noisy conditioning, where small perturbations in text or multimodal embeddings can cascade over timesteps and cause semantic drift. Existing corruption strategies from image diffusion (Gaussian, Uniform) fail in video settings because static noise disrupts temporal fidelity. In this paper, we propose CAT-LVDM, a corruption-aware training framework with structured, data-aligned noise injection tailored for video diffusion. Our two operators, Batch-Centered Noise Injection (BCNI) and Spectrum-Aware Contextual Noise (SACN), align perturbations with batch semantics or spectral dynamics to preserve coherence. CAT-LVDM yields substantial gains: BCNI reduces FVD by 31.9 percent on WebVid-2M, MSR-VTT, and MSVD, while SACN improves UCF-101 by 12.3 percent, outperforming Gaussian, Uniform, and even large diffusion baselines like DEMO (2.3B) and Lavie (3B) despite training on 5x less data. Ablations confirm the unique value of low-rank, data-aligned noise, and theory establishes why these operators tighten robustness and generalization bounds. CAT-LVDM thus sets a new framework for robust video diffusion, and our experiments show that it can also be extended to autoregressive generation and multimodal video understanding LLMs. Code, models, and samples are available at https://github.com/chikap421/catlvdm

Prefix-sharing among multiple prompts presents opportunities to combine the operations of the shared prefix, while attention computation in the decode stage, which becomes a critical bottleneck with increasing context lengths, is a memory-intensive process requiring heavy memory access on the key-value (KV) cache of the prefixes. Therefore, in this paper, we explore the potential of prefix-sharing in the attention computation of the decode stage. However, the tree structure of the prefix-sharing mechanism presents significant challenges for attention computation in efficiently processing shared KV cache access patterns while managing complex dependencies and balancing irregular workloads. To address the above challenges, we propose a dedicated attention kernel to combine the memory access of shared prefixes in the decoding stage, namely CoDec. CoDec delivers two key innovations: a novel shared-prefix attention kernel that optimizes memory hierarchy and exploits both intra-block and inter-block parallelism, and a comprehensive workload balancing mechanism that efficiently estimates cost, divides tasks, and schedules execution. Experimental results show that CoDec achieves an average $1.9\times$ speedup and $120.9\times$ memory access reduction compared to the state-of-the-art FlashDecoding kernel regarding attention computation in the decode stage and $3.8\times$ end-to-end time per output token compared to the vLLM.