Retrieval-Augmented Generation (RAG) has become a foundational paradigm for integrating AI agents with external knowledge. However, current RAG models remain largely constrained by static retrieval pipelines and limited capacity for adaptive reasoning over hierarchical knowledge structures. As AI agents increasingly operate in dynamic, information-rich environments, there is a growing need for models that can reason across modalities while continuously evolving their knowledge representations. We introduce HFlow, a self-evolving multimodal agentic RAG framework grounded in hypergraph representations and hyperbolic flow-based reasoning. In our formulation, heterogeneous modalities, including text, images, audio, and structured data, are modelled as nodes within a multimodal hypergraph, while hyperedges capture higher-order semantic and cross-modal relationships. By embedding this structure in hyperbolic space, the framework preserves hierarchical and compositional knowledge. This work proposes a shift from static retrieval to continuous knowledge navigation, where reasoning emerges through geometry-aware traversal of multimodal knowledge manifolds. The proposed framework unifies retrieval, reasoning, and adaptation within a single agentic architecture, offering a new direction for scalable, context-aware AI models. We discuss early empirical evidence demonstrating improved robustness and reasoning flexibility compared to conventional Euclidean and pipeline-based multimodal RAG approaches and outline future opportunities for self-improving knowledge agents.

Research on Information Retrieval (IR) has historically prioritised high-resource languages such as English and Chinese, with less attention given to many low-resource languages. For example, Shona, a Bantu language spoken by approximately 12 million people in Zimbabwe and neighbouring countries, remains under-explored in IR research despite its widespread societal use in Southern Africa. In this work, we present a preliminary study of Shona IR using sparse and dense retrieval models, demonstrating significant performance limitations due to morphological complexity and data scarcity. Based on these findings, we propose to develop a framework to advance Shona IR by developing a large-scale benchmark dataset to support morphology-aware retrieval. We hypothesise that improving Shona IR supports equitable access to digital information and enables language-inclusive AI technologies aligned with global development priorities such as accessibility to education, dissemination of healthcare information, and digital inclusion.

\beginabstract Recently, modelling heterogeneous interactions in recommender systems has attracted research interest. Real-world scenarios involve sequential multi-type user-item interactions such as ''shape view'', ''shape add-to-favourites'', ''shape add-to-cart'' and ''shape purchase''. Graph Neural Network (GNN) methods have been widely adopted in Representation Learning of similar sequential user-item interactions. Promising results have been achieved by the integration of GNNs and transformers for self-attention. However, GNN based methods suffer from limited capability in handling global user-item interaction dependencies, particularly for long sequences. Moreover, these models require high computational cost of transformers, due to the quadratic memory and time complexity with respect to sequence length. This results in memory bottlenecks and slow training especially in computational resource-constrained environments. To address these challenges, we propose the FATH model which employs Flash Attention mechanism to reduce the high-bandwidth memory usage over higher-order user-item interaction sequences. Experimental results show that our model improves the training speed and reduces the memory usage with better recommendation performance in comparison with the state-of the art baselines.

In this work, we evaluate contrastive models for the task of image retrieval. We hypothesise that models that are learned to encode semantic similarity among instances via discriminative learning should perform well on the task of image retrieval, where relevancy is defined in terms of instances of the same object. Through our extensive evaluation, we find that representations from models trained using contrastive methods perform on-par with (and outperforms) a pre-trained supervised baseline trained on the ImageNet labels in retrieval tasks under various configurations. This is remarkable given that the contrastive models require no explicit supervision. Thus, we conclude that these models can be used to bootstrap base models to build more robust image retrieval engines.

This work proposes a simple instance retrieval pipeline based on encoding the convolutional features of CNN using the bag of words aggregation scheme (BoW). Assigning each local array of activations in a convolutional layer to a visual word produces an assignment map, a compact representation that relates regions of an image with a visual word. We use the assignment map for fast spatial reranking, obtaining object localizations that are used for query expansion. We demonstrate the suitability of the BoW representation based on local CNN features for instance retrieval, achieving competitive performance on the Oxford and Paris buildings benchmarks. We show that our proposed system for CNN feature aggregation with BoW outperforms state-of-the-art techniques using sum pooling at a subset of the challenging TRECVid INS benchmark.

This paper explores the potential for using Brain Computer Interfaces (BCI) as a relevance feedback mechanism in content-based image retrieval. Several experiments are performed using a rapid serial visual presentation (RSVP) of images at different rates (5Hz and 10Hz) on 8 users with different degrees of familiarization with BCI and the dataset. We compare the feedback from the BCI and mouse-based interfaces in a subset of TRECVid images, finding that, when users have limited time to annotate the images, both interfaces are comparable in performance. Comparing our best users in a retrieval task, we found that EEG-based relevance feedback can outperform mouse-based feedback.

This demonstration shows the integration of video analysis and search tools to facilitate the interactive retrieval of video segments depicting specific activities from surveillance footage. The implementation was developed by members of the SAVASA project for participation in the interactive surveillance event detection (SED) task of TRECVid 2012. This year, for the first time, the purpose of the interactive SED task was to evaluate systems' ability to support users in identifying video segments that depict a specific activity (event) in a large collection of surveillance video footage. Project partners worked together to analyse video and provide a query interface enabling users to search and identify matching video segments. The collaborative integration of components from multiple partners and the participation of end user partners in evaluating the system are the novel aspects of this work.

This paper presents work on integrating multiple computer vision-based approaches to surveillance video analysis to support user retrieval of video segments showing human activities. Applied computer vision using real-world surveillance video data is an extremely challenging research problem, independently of any information retrieval (IR) issues. Here we describe the issues faced in developing both generic and specific analysis tools and how they were integrated for use in the new TRECVid interactive surveillance event detection task. We present an interaction paradigm and discuss the outcomes from face-to-face end user trials and the resulting feedback on the system from both professionals, who manage surveillance video, and computer vision or machine learning experts. We propose an information retrieval approach to finding events in surveillance video rather than solely relying on traditional annotation using specifically trained classifiers.

We demonstrate a multimedia content information retrieval engine developed for audiovisual digital libraries targeted at media professionals. It is the first of three multimedia IR systems being developed by the AXES project. The system brings together traditional text IR and state-of-the-art content indexing and retrieval technologies to allow users to search and browse digital libraries in novel ways. Key features include: metadata and ASR search and filtering, on-the-fly visual concept classification (categories, faces, places, and logos), and similarity search (instances and faces).

In this paper we describe the design considerations for a touchscreen visual lifelog browser. Visual lifelogs are large collections of photographs which represent a person's experiences. Lifelogging devices, such as the wearable camera known as SenseCam, can record thousands of images per day. Utilizing the approach of event segmentation to organize and present these images, we have designed an interface to present lifelog collections for touchscreen interaction, thus increasing accessibility for users.

The paper reports a new fast intra prediction algorithms based on separating the transformed coefficients of neighboring blocks. The prediction blocks are obtained from the transformed and quantized neighboring blocks that generate minimum distortion for each DC and AC coefficients. To obtain fast coding with comparable coding efficiency compared to H.264/AVC, we present the full block search prediction (FBSP) and the edge based distance prediction (EBDP). These are immune to both the intra prediction error and the drift propagation; in addition, do not require a low pass filtering named extrapolation and mode decisions to obtain a prediction block. In FBSP, eight neighbor blocks (B(i-j)) are chosen to be candidates by observing the edge direction of the current block ((0,0)), where (i,j) are relative displacement position to the current block denoted as (i,j) G {(0,-1), (-1, 0), (-1, -1), (+1, -1), (-2, -1), (-1, -2), (+2, -1), (+1, -2)}, and the prediction block can be obtained by combining DC and AC coefficients of a neighboring block generating minimum distortion of residues. In EBSP, the edge directions of the current block are estimated from DCT coefficients, and intuitionally the prediction block is selected by observing the minimum edge distance (difference) compared to the current block without any distortion measurements. We have applied the proposed algorithms to MPEG test sequences which are Foreman, Hall, Mother and Daughter and Mobile. A PSNR (peak signal to noise ratio) gain of -2 to +0.8dB is observed at various bit rates shown in Figure l(a)(b), where + indicates PSNR degradation, - denotes PSNR gain compared to H.264/AVC intra coding. In case of a complex scene in terms of edges, the prediction error could be more dominant rather than one caused by quantization, the proposed method shows the same or better coding efficiency due to its insensitivity of the prediction error and the drift propagation. Analyzing of computational complexity, the proposed algorithms require less than 15% complexity compared to low complexity mode of H.264/AVC in Figure 1(c) since they does not perform an extrapolation and mode decisions which generates intensive computational complexity in H.264/AVC intra coding.

In this paper, we evaluate the appearance tracking performance of multiple fusion schemes that combine information from standard CCTV and thermal infrared spectrum video for the tracking of surveillance objects, such as people, faces, bicycles and vehicles. We show results on numerous real world multimodal surveillance sequences, tracking challenging objects whose appearance changes rapidly. Based on these results we can determine the most promising fusion schemes

Technological developments in content-based analysis of digital video information are undergoing much progress, with ideas for fully automatic systems now being proposed and demonstrated. Yet because we do not yet have robust operational video retrieval systems that can be deployed and used, the usual HCI practise of conducting a usage study and an informed iterative system design is thus not possible. Físchlár-News is one of the first automatic, content-based broadcast news analysis and archival systems that process broadcast news video so that users can search, browse, and play it in an easy-to-use manner with a conventional web browser. The system incorporates a number of state-of-the-art research components, some of which are not yet considered mature technology, yet it has been built to be robust enough to be deployed to users who are interested in access to daily news throughout a university campus. In this article we report and discuss a user-evaluation study conducted with 16 users, each of whom utilized the system freely for a one month period. Results from a detailed qualitative analysis are presented, looking at collected questionnaires, incident diaries, and interaction-log data. The findings suggest that our users employed the system in conjunction with their other news update methods, such as watching TV news at home and browsing online news websites at their workplace, their major concerns being up-to-dateness and coverage of the news content. They tried to accommodate the system to fit their established web browsing habits, and they found local news content and the ability to play self-contained news stories on their desktop as major values of the system. Our study also resulted in a detailed wishlist of new features which will help in the further development of both our and others' systems.