Language models (LMs) are trained to excel at predicting the next word in the sequence given prior context, and humans also share this predictability in reading comprehension. Neuroscience research reveals that next-word predictability influences brain response, as recorded at millisecond resolution using electroencephalography (EEG). While our evidence indicates that advanced LMs achieve accuracies closely aligned with human performance at the next-word prediction task, this raises the question: Does higher prediction accuracy necessarily mean that these models adequately capture the cognitive signals associated with human reading comprehension? Here, we generate regressors for both humans and LMs based on two information measures, including top-1 prediction and surprisal, to predict event-related potential (ERP) elicited from EEG recordings which reflect different stages of cognitive processing during reading. We argue that modelling ERP patterns offers fine-grained analysis of the cognitive plausibility of various LMs during reading. Our results indicate that only surprisal potentially correlates with language-processing ERPs, especially for open-class words with high semantic content. Moreover, our findings challenge the assumption that scaling LMs with increased parameters and computational budgets will consistently lead to improved convergence with human-like linguistic processing.

Vision-Language Models (VLMs) are typically pre-trained on large-scale image-text datasets to capture semantic correspondences between visual content and natural language. However, they remain surprisingly brittle to negation: models often rely on shallow word co-occurrence and are easily distracted by misleading or irrelevant textual cues, even when their overall retrieval or classification performance is strong. Moreover, directly finetuning on negation data can interfere with previously acquired knowledge, causing noticeable degradation on standard vision-language benchmarks. To tackle these issues, this work introduces HANCLIP (Hyperbolic + Angular + Negation), a family of VLMs that explicitly restructures the embedding space to encode "what an image is not" alongside "what it is." HANCLIP is trained on a compact set of 20,000 image-text quadruplets and combines a hyperbolic formulation, which models hierarchical semantic relations and asymmetries, with an angular triplet objective that drives systematic separation between negated descriptions and their corresponding positives. This geometry-aware design strengthens negation sensitivity while preserving the global structure of pretrained representations, rather than overwriting them. Extensive experiments across multiple vision-language tasks show that HANCLIP delivers consistent gains on the negation-focused NegBench benchmark, while maintaining competitive or improved performance on standard classification and image-text retrieval benchmarks. The framework is model-agnostic and can be plugged into CLIP, LongCLIP, SmartCLIP, and HiMo-CLIP without large-scale retraining, demonstrating that a carefully designed geometric objective can substantially extend the reasoning capabilities of existing VLMs using only modest additional data.

Large-scale video collections require retrieval systems that understand both semantic content and temporal structure. While vision–language models have improved cross-modal retrieval, searching long-form videos remains challenging, especially for abstract queries and event sequences. This paper presents an overview of the Ho Chi Minh City AI Challenge 2025, a large-scale evaluation campaign inspired by the Video Browser Showdown and the Lifelog Search Challenge. The challenge includes multiple query types: Textual and Visual Known-Item Search and Question Answering. It also introduces a new task, Temporal Retrieval and Alignment of Key Events (TRAKE), which involves retrieving a relevant video and aligning multiple key events in time. Progressive hints are provided to simulate realistic query refinement. We describe the dataset, task design, evaluation protocol, and analyze team performance and retrieval behavior. The results highlight both recent progress and remaining challenges in temporally-aware video retrieval.

Image-Guided Retrieval with Optional Text (IGROT) unifies visual retrieval (without text) and composed retrieval (with text). Despite its relevance in applications like Google Image and Bing, progress has been limited by the lack of an accessible benchmark and methods that balance performance across subtasks. Large-scale datasets such as MagicLens are comprehensive but computationally prohibitive, while existing models often favor either visual or compositional queries. We introduce FIGROTD, a lightweight yet high-quality IGROT dataset with 16,474 training triplets and 1,262 test triplets across CIR, SBIR, and CSTBIR. To reduce redundancy, we propose the Variance Guided Feature Mask (VaGFeM), which selectively enhances discriminative dimensions based on variance statistics. We further adopt a dual-loss design (InfoNCE + Triplet) to improve compositional reasoning. Trained on FIGROTD, VaGFeM achieves competitive results on nine benchmarks, reaching 34.8 mAP@10 on CIRCO and 75.7 mAP@200 on Sketchy, outperforming stronger baselines despite fewer triplets. Our dataset and code are available at here .

Modern Video Retrieval systems face challenges in computational efficiency and semantic depth when handling complex queries, particularly those with time-sensitive requirements. These systems typically rely on a “flat” index structure that encodes each frame independently, resulting in high search costs and difficulty capturing higher-level events or context semantics. To address these limitations, we propose a novel three-level hierarchical index concept that organizes video data at different semantic abstraction levels. The first level involves embedding vectors for individual frames to facilitate fine-grained retrieval. The second level groups visually similar frames into “shots” and encodes them into a semantic temporal representation. The top layer uses a Visual-Language Model (VLM) to identify and group frames related to narrative actions. This architecture allows the system to first quickly identify high-level related scenes or actions, and then refine the results by searching within individual frames within those groups. Our approach helps users to query data at the most relevant conceptual level.

Multimedia research is framed through algorithms, datasets, and systems, but at its heart lies content that is deeply human. Few forms of content illustrate this better than classical music. Long before music becomes data to be recorded, generated, searched, or retrieved, it is imagined by composers and brought to life by performers. At ACM Multimedia 2025 in Dublin, this human origin of multimedia took centre stage in a unique social event that bridged classical music and multimedia content analysis. This event was the fourth supported by ACM SIGMM in the framework of Music Meets Science program (at CBMI'2022, CBMI'2023, CBMI'2024).

We introduce OpenLifelogQA, a large-scale open-ended lifelog QA dataset constructed from 18 months of multimodal lifelog data. Lifelogging is the passive collection and analysis of personal daily activities using wearable devices, producing rich multimodal data such as images, locations, and biometrics. Question answering (QA) over lifelog data enables users to interactively query their own experiences, supporting applications in memory support, lifestyle analysis, and personal assistance. OpenLifelogQA contains 14,187 Q A pairs spanning multiple question types and difficulty levels, designed to support robust evaluation in realistic settings. Compared with prior resources, OpenLifelogQA offers greater diversity and practicality for real-world applications. To establish baselines, we evaluate the LLaVA-NeXT-Interleave 7B model, achieving 89.7

Efficient video discovery and retrieval remains a significant challenge, especially in massive video archives that can contain thousands of hours of content. Traditional browsing methods, such as timeline previews or metadata-based searches, only allow users to access content sequentially. This makes it difficult to get a comprehensive view or quickly identify the required content. Inspired by the concept of Video Tapestry, we present TapesVRy, a system that enables interactive, immersive video discovery. Instead of presenting frames in a flat layout, our system arranges and blends keyframes and short clips into a 360 ^∘ panoramic tapestry that users can explore with Meta Quest VR headsets. Videos are automatically clustered by semantic similarity using multimodal AI models, grouping related videos into thematic “universes”. In each universe, users can freely navigate the 360 ^∘ space, zooming in or out to adjust the density of the content. By directly interacting with the image areas, they can immediately access the original video segments. This approach transforms passive video viewing into an active exploration process, allowing users to browse and intuitively grasp multiple videos in a short time.

Long-horizon video editing agents receive final-product feedback only after many interdependent decisions. Yet editing quality is subjective, admits multiple valid solutions, and is not meaningfully calibrated across heterogeneous requests, making a global scalar objective both ambiguous and temporally uninformative. Our key observation is that fixing the request, materials, and production constraints converts this subjective objective into an ordinal comparison among directly comparable alternatives. We introduce Group-Relative Preference Backpropagation (GRPB), which transforms same-task rankings into zero-sum advantages and redistributes them as bounded credit over semantic editing segments. A lagged allocator and guarded transmission prevent current judgments or unreliable estimates from directly shaping the same rollout group. We manually construct a project-disjoint, horizon-stratified suite of realistic editing tasks for training and controlled evaluation. Across matched baselines, credit interventions, external benchmarking, and blinded human evaluation, GRPB improves both editing behavior and rendered products. The resulting 9B Crayotter model surpasses several proprietary systems on AgenticVBench, supporting task-local preference reduction as a practical approach to learning from subjective, delayed outcomes. Code and all supporting materials are publicly available at https://github.com/idwts/Crayotter.

The Video Browser Showdown (VBS) is an international competition evaluating systems for interactive video retrieval across large datasets. In VBS 2026, we present SnapMind, the next generation of SnapSeek, which introduces an LLM-based Planner acting as an AI agent to support both expert and novice users. Given a query and the registry of retrieval components (e.g., text, image, OCR, ADL, object), the Planner generates candidate execution plans describing how these components can be combined. Users may select, edit, or discard plans, while the system executes them incrementally with normalization and fusion strategies. By embedding this AI agent into the loop, the system demonstrates agentic AI behavior: it can propose strategies, adapt to feedback, with the user controlling its participation via three autonomy levels. This design helps novices transform vague ideas into executable retrieval strategies while allowing experts to retain fine-grained control.

Progress in gaze-informed multimedia retrieval has been constrained by the absence of datasets that capture visual attention in realistic, goal-driven search scenarios. To address this, we present a new eye-tracking dataset collected from 24 participants performing five goal-driven tasks using two multimedia collections: the Vimeo Creative Commons Collection (V3C) and ImageNet. Tasks include ad-hoc exploration, known-item exploration, textual ranking, visual ranking, and interactive retrieval. In total, the dataset comprises over 6 h of gaze recordings, amounting to approximately 1.4 million gaze points, along with synchronized interaction logs and detailed task metadata. This multimodal resource enables fine-grained analysis of visual attention, user strategies, and search behavior across diverse task types and multimedia domains. Potential applications include gaze-informed content ranking, task-aware saliency prediction, media quality assessment, and cross-dataset benchmarking. By making this dataset available at https://osf.io/rzp64 , we aim to provide a reproducible, versatile resource that advances the study of user-adaptive multimedia retrieval.

This report summarizes the contributions and results of the 1st Asynchronous CASTLE Challenge at the Joint Egocentric Vision Workshop in conjunction with CVPR 2026.

No-Reference Image Quality Assessment (NR-IQA) remains a challenging task due to the need to model the complex interplay between global semantic content and local distortion artifacts, central to human visual perception. Existing methods often rely on hand-crafted features, explicit feature separation, or multi-stage restoration pipelines, which fail to adaptively integrate cross-scale information. To address these limitations, we propose DTIQA, an end-to-end framework named Dual-Path Transformer for Image Quality Assessment. DTIQA introduces a Global–Local Gated Decomposition (GLGD) module to generate complementary content-aware and distortion-aware feature representations without explicit feature separation. A bidirectional cross-scale attention mechanism further refines these features, enabling adaptive convergence of contextual importance and degradation evidence. Extensive experiments on eight benchmark IQA datasets demonstrate that DTIQA achieves state-of-the-art performance, outperforming competing methods with strong generalization and stable training dynamics. The official implementation is available at https://github.com/algaradi/DTIQA (DOI: https://doi.org/10.5281/zenodo.18842363 ).

Humans invented reading and have passed down this complex skill across generations through language. This study provides empirical evidence of the neural mechanisms underlying bottom-up (related to high-order linguistic structure) and top-down (related to next-word predictability) processes, which interact to guide comprehension during reading. While previous studies have focused on either the N400 effects of predictability or lexical categories, research on how predictability influences N400 responses across different lexical categories is limited, mainly due to constraints in publicly available datasets. Here, we examine how predictability influences brain responses, recorded at millisecond resolution using electroencephalography (EEG), with a focus on the N400 time window (300-500 ms post-stimulus) across different lexical and grammatical categories. Our results indicate that significant differences in N400 responses between high and low cloze probability levels were more pronounced for content words than function words. Among the two primary content categories, verbs exhibited greater N400 differences than nouns, while nouns carried more distinct information about their predictability than verbs. Moreover, we demonstrate that the decoding technique is more effective than the event-related potential (ERP) traditional analysis in capturing more detailed and distinct representations of cognitive processes over time.

In this work, we focus on assisting users in finding information that may have been unintentionally overlooked. Our system supports not only experienced users but also newcomers to Video Browser Showdown systems, enabling them to search for information more quickly and accurately. During the querying process, users might unintentionally miss important images, including those they are specifically looking for. By leveraging eye-tracking technology, our system records the user's gaze duration on each image and optimises the presentation of potentially relevant images, which should lead to better performance when using interactive systems under time-pressure.