LLM-powered search agents are increasingly being used for multi-step information seeking tasks, yet the IR community lacks empirical understanding of how agentic search sessions unfold and how retrieved evidence is reflected in later queries. This paper presents a large-scale log analysis of agentic search based on 14.44M search requests (3.97M sessions) collected from DeepResearchGym, i.e., an open-source search API accessed by external agentic clients. We sessionize the logs, assign session-level intents and step-wise query-reformulation labels using LLM-based annotation, and propose Context-driven Term Adoption Rate (CTAR) to quantify whether newly introduced query terms are lexically traceable to previously retrieved evidence. Our analyses reveal distinctive behavioral patterns. First, over 90\% of multi-turn sessions contain at most ten steps, and 89\% of inter-step intervals fall under one minute. Second, behavior varies by intent. Fact-seeking sessions exhibit high repetition that increases over time, while sessions requiring reasoning sustain broader exploration. Third, query reformulations are often traceable to retrieved evidence across steps. On average, 54\% of newly introduced query terms appear in the accumulated evidence context, with additional traceability to earlier steps beyond the most recent retrieval. These findings provide candidate signals for repetition-aware stopping, intent-adaptive retrieval budgeting, and explicit cross-step context tracking. We released the anonymized logs, making them available at a public HuggingFace~\chref{https://huggingface.co/datasets/cx-cmu/deepresearchgym-agentic-search-logs}{repository}.

Neural retrieval models excel in Web search, but their training requires substantial amounts of labeled query-document pairs, which are costly to obtain. With the widespread availability of Web document collections like ClueWeb22, synthetic queries generated by large language models offer a scalable alternative. Still, synthetic training queries often vary in quality, which leads to suboptimal downstream retrieval performance. Existing methods typically filter out noisy query-document pairs based on signals from an external re-ranker. In contrast, we propose a framework that leverages Direct Preference Optimization (DPO) to integrate ranking signals into the query generation process, aiming to directly optimize the model towards generating high-quality queries that maximize downstream retrieval effectiveness. Experiments show higher ranker-assessed relevance between query-document pairs after DPO, leading to stronger downstream performance on the MS~MARCO benchmark when compared to baseline models trained with synthetic data.

Methods based on Contrastive Language-Image Pre-training (CLIP) are nowadays extensively used in support of vision-and-language tasks involving remote sensing data, such as cross-modal retrieval. The adaptation of CLIP to this specific domain has relied on model fine-tuning with the standard contrastive objective, using existing human-labeled image-caption datasets, or using synthetic data corresponding to image-caption pairs derived from other annotations over remote sensing images (e.g., object classes). The use of different pre-training mechanisms has received less attention, and only a few exceptions have considered multilingual inputs. This work proposes a novel vision-and-language model for the remote sensing domain, exploring the fine-tuning of a multilingual CLIP model and testing the use of a self-supervised method based on aligning local and global representations from individual input images, together with the standard CLIP objective. Model training relied on assembling pre-existing datasets of remote sensing images paired with English captions, followed by the use of automated machine translation into nine additional languages. We show that translated data is indeed helpful, e.g. improving performance also on English. Our resulting model, which we named Remote Sensing Multilingual CLIP (RS-M-CLIP), obtains state-of-the-art results in a variety of vision-and-language tasks, including cross-modal and multilingual image-text retrieval, or zero-shot image classification.

Predicting the success of Conversational Task Assistants (CTA) can be critical to understand user behavior and act accordingly. In this paper, we propose TB-Rater, a Transformer model which combines conversational-flow features with user behavior features for predicting user ratings in a CTA scenario. In particular, we use real human-agent conversations and ratings collected in the Alexa TaskBot challenge, a novel multimodal and multi-turn conversational context. Our results show the advantages of modeling both the conversational-flow and behavioral aspects of the conversation in a single model for offline rating prediction. Additionally, an analysis of the CTA-specific behavioral features brings insights into this setting and can be used to bootstrap future systems.

We present a method for performing zero-shot Dialogue State Tracking (DST) by casting the task as a learning-to-ask-questions framework. The framework learns to pair the best question generation (QG) strategy with in-domain question answering (QA) methods to extract slot values from a dialogue without any human intervention. A novel self-supervised QA pretraining step using in-domain data is essential to learn the structure without requiring any slot-filling annotations. Moreover, we show that QG methods need to be aligned with the same grammatical person used in the dialogue. Empirical evaluation on the MultiWOZ 2.1 dataset demonstrates that our approach, when used alongside robust QA models, outperforms existing zero-shot methods in the challenging task of zero-shot cross domain adaptation-given a comparable amount of domain knowledge during data creation. Finally, we analyze the impact of the types of questions used, and demonstrate that the algorithmic approach outperforms template-based question generation.

People often ask questions about places, and this is reflected on the frequency of geo-spatial queries made to information retrieval and question answering systems. Recent developments associated to these two types of systems rely on deep neural networks, specifically on methods for passage retrieval based on Transformer models, trained on large datasets like MS-MARCO. Despite significant progress in approaches for retrieving (or re-ranking) passages from a document collection according to their relevance to an input query, few studies have specifically looked at geo-spatial queries (i.e., where-questions directly concerning locations, and also questions covering other informational needs relating to places, their types, and affordances). In this work, we explore neural retrieval models in the context of geo-spatial queries, using a subset of MS-MARCO with questions and passages containing place-names. After characterizing the subset of MS-MARCO, we analyzed a re-ranking strategy based on geographic distance, which we argue to be useful for selecting hard negative examples for model training. Then, we fine-tuned neural ranking models, following bi-encoder or cross-encoder strategies, using the MS-MARCO subset together with a geographically-aware negative sampling procedure. Experimental results show that the fine-tuned models can indeed achieve a superior performance. We also describe a simple knowledge distillation procedure to further improve the computationally more efficient bi-encoder models, using the results of the cross-encoder.

Media editors in the newsroom are constantly pressed to provide a"like-being there" coverage of live events. Social media provides a disorganised collection of images and videos that media professionals need to grasp before publishing their latest news updated. Automated news visual storyline editing with social media content can be very challenging, as it not only entails the task of finding the right content but also making sure that news content evolves coherently over time. To tackle these issues, this paper proposes a benchmark for assessing social media visual storylines. The SocialStories benchmark, comprised by total of 40 curated stories covering sports and cultural events, provides the experimental setup and introduces novel quantitative metrics to perform a rigorous evaluation of visual storytelling with social media data.

Distributing multimedia indexes to multiple nodes enables search over very large datasets (i.e., over one billion images and videos), but comes with a set of challenges: \textithow to distribute documents and queries effectively across nodes to support concurrent querying? andhow to deal with the increased potential for lack of response from nodes (e.g., node fail-stops or dropping of network packages)? An index where partitions are based on the distribution of feature vectors in the original space can improve redundancy and increase efficiency: nearest neighbors are only present on a small, set number of partitions, reducing the number of nodes to inspect for each query. This paper describes how sparse hashes can help find this balance and create better distribution policies for high-dimensional feature vectors. Inspired by existing literature on distributed text and media indexes, our proposal distributes and balances documents and queries to a subset of the nodes, according to their orthogonal similarities. We performed exhaustive benchmarks of our approach on a commercial cloud service. Experiments on a one billion vector dataset show that our approach has a low partitioning overhead (3 to 5 ms per query), achieves balanced document and query distribution (the variation in document and query distribution across nodes is smaller than 1% and 10%, respectively), handles concurrent queries effectively and degrades gracefully with node failures (less than 2% of precision loss per node down).

Newsworthy events are broadcast through multiple mediums and prompt the crowds to produce comments on social media. In this paper, we propose to leverage on this behavioral dynamics to estimate the most relevant time periods for an event (i.e., query). Recent advances have shown how to improve the estimation of the temporal relevance of such topics. In this approach, we build on two major novelties. First, we mine temporal evidences from hundreds of external sources into topic-based external collections to improve the robustness of the detection of relevant time periods. Second, we propose a formal retrieval model that generalizes the use of the temporal dimension across different aspects of the retrieval process. In particular, we show that temporal evidence of external collections can be used to (i) infer a topic's temporal relevance, (ii) select the query expansion terms, and (iii) re-rank the final results for improved precision. Experiments with TREC Microblog collections show that the proposed time-aware retrieval model makes an effective and extensive use of the temporal dimension to improve search results over the most recent temporal models. Interestingly, we observe a strong correlation between precision and the temporal distribution of retrieved and relevant documents.

Combining multiple retrieval functions can lead to notable gains in retrieval performance. Learning to Rank (LETOR) techniques achieve outstanding retrieval results, by learning models with no bounds on model complexity. Often, minor retrieval gains are attained at a significant cost in model complexity. This paper focuses on the research question:can less complex models achieve results comparable to LETOR models? In this paper, we investigate an approach for the selection and fusion of rank lists with low-complexity models. The described Learning to Fuse (L2F) algorithm, is a supervised rank fusion procedure that controls the model complexity by discarding rank lists that bring minor improvements to final rank. Evaluation results, on two different datasets, show that it is indeed possible to achieve a retrieval performance comparable to LETOR methods, using only 3-5% of the rank lists of the number of rank lists used by LETOR methods.

News editors need to find the photos that best illustrate a news piece and fulfill news-media quality standards, while being pressed to also find the most recent photos of live events. Recently, it became common to use social-media content in the context of news media for its unique value in terms of immediacy and quality. Consequently, the amount of images to be considered and filtered through is now too much to be handled by a person. To aid the news editor in this process, we propose a framework designed to deliver high-quality, news-press type photos to the user. The framework, composed of two parts, is based on a ranking algorithm tuned to rank professional media highly and a visual SPAM detection module designed to filter-out low-quality media. The core ranking algorithm is leveraged by aesthetic, social and deep-learning semantic features. Evaluation showed that the proposed framework is effective at finding high-quality photos (true-positive rate) achieving a retrieval MAP of 64.5% and a classification precision of 70%.

3D video is introducing great changes in many health related areas. The realism of such information provides health professionals with strong evidence analysis tools to facilitate clinical decision processes. Speech and language therapy aims to help subjects in correcting several disorders. The assessment of the patient by the speech and language therapist (SLT), requires several visual and audio analysis procedures that can interfere with the patient's production of speech. In this context, the main contribution of this paper is a 3D video system to improve health information management processes in speech and language therapy. The 3D video retrieval and management system supports multimodal health records and provides the SLTs with tools to support their work in many ways: (i) it allows SLTs to easily maintain a database of patients' orofacial and speech exercises; (ii) supports three-dimensional orofacial measurement and analysis in a non-intrusive way; and (iii) search patient speech-exercises by similar facial characteristics, using facial image analysis techniques. The second contribution is a dataset with 3D videos of patients performing orofacial speech exercises. The whole system was evaluated successfully in a user study involving 22 SLTs. The user study illustrated the importance of the retrieval by similar orofacial speech exercise.

This paper addresses the problem of balanced, redundant indexing of media information. Our goal is to partition and distribute the search index, taking advantage of the distributed systems properties: balanced load across nodes, redundancy on node down and efficient node usage under concurrent querying. We follow an information compression approach to solve this problem and propose to represent data with overcomplete codebooks, where each document is represented by only a few codewords and an indexing node is responsible for several codewords. Quantization algorithms are designed to fit the original data as best as possible, leading to bias towards codewords that fit the principal directions of data. In this paper, we propose the balanced KSVD (B-KSVD) algorithm, that distributes the allocation of data across a balanced number of codewords, according to the global distribution of data. Indexing experiments showed that B-KSVD can achieve 38% 1-recall by inspecting only 1% of the full index, distributed over 10 partitions. Traditional methods based on k-means need to either use larger codebooks or to inspect a larger portion of the index to achieve the same retrieval performance.

In Twitter, and other microblogging services, the generation of new content by the crowd is often biased towards immediacy: what is happening now. Prompted by the propagation of commentary and information through multiple mediums, users on the Web interact with and produce new posts about newsworthy topics and give rise to trending topics. This paper proposes to leverage on the behavioral dynamics of users to estimate the most relevant time periods for a topic. Our hypothesis stems from the fact that when a real-world event occurs it usually has peak times on the Web: a higher volume of tweets, new visits and edits to related Wikipedia articles, and news published about the event.

In this paper we propose a high-dimensional indexing technique, based on sparse approximation techniques to speed up the search and retrieval of similar images given a query image feature vector. Feature vectors are stored on an inverted indexed based on a sparsifying dictionary for l0 regression, optimized to reduce the data dimensionality. It concentrates the energy of the original vector on a few coefficients of a higher dimensional representation. The index explores the coefficient locality of the sparse representations, to guide the search through the inverted index. Evaluation on three large-scale datasets showed that our method compares favorably to the state-of-the-art. On a 1 million dataset of SIFT vectors, our method achieved 60.8% precision at 50 by inspecting only 5% of the full dataset, and by using only 1/4 of the time a linear search takes.

Multimedia search systems must deal with an increasingly large and heterogeneous amount of data. Several challenges exist when deploying real-world search engines for such data. Existing literature does not properly tackle the many efficiency issues that such task requires. In this paper, we address several of the key efficiency aspects required to deploy a distributed search engine, capable of handling several millions of multimedia documents. The search engine builds on a framework designed to: first, ease the distribution of documents and queries across cluster-nodes, second, index media efficiently for fast similarity search and third aggregate ranked results from several heterogeneous sources. Moreover, the proposed framework is flexible enough to support several state-of-the-art indexing and aggregation techniques.

Reputation analysis is naturally linked to a sentiment analysis task of the targeted entities. This analysis leverages on a sentiment lexicon that includes general sentiment words and domain specific jargon. However, in most cases target entities are themselves part of the sentiment lexicon, creating a loop from which it is difficult to infer an entity reputation. Sometimes, the entity became a reference in the domain and is vastly cited as an example of a highly reputable entity. For example, in the movies domain it is not uncommon to see reviews citing Batman or Anthony Hopkins as esteemed references. In this paper we describe an unsupervised method for performing a simultaneous-analysis of the reputation of multiple named-entities. Our method jointly extracts named entities reputation and a domain specific sentiment lexicon. The objective is two-fold: (1) named-entities are naturally ranked by our method and (2) we can build a reputation graph of the domain's named entities. This framework has immediate applications in terms of visualization or search by reputation.

In this paper, we propose a framework to characterize and compare two search engine results. Typical user-queries are ambiguous and, consequentially, each search engine will compute ranks in different manners, attempting to answer them in the best possible way. Thus, each search engine will have its own bias. Given the importance of the first page results in Web Search Engines, in this paper we propose a framework to assess the information presented in the first page by measuring the information entropy and the correlations between two ranks. Employing the recently proposed Rank-Biased Overlap measure [2] we compare to which extent do Bing and Google rankings in fact differ. We also extend this measure and propose a measure for comparing the information entropy present in two ranks. The proposed measure is based on the correlation of two ranks and the application of Jensen-Shannon's divergence among two document sets. Our methodology starts with 40,000 user queries and crawls the search results for these queries on both search engines. The results allow us to determine the search engines correlations, crawling coverage, information overlap, and information entropy.

This article is set in the context of searching text and image repositories by keyword. We develop a unified probabilistic framework for text, image, and combined text and image retrieval that is based on the detection of keywords (concepts) using automated image annotation technology. Our framework is deeply rooted in information theory and lends itself to use with other media types. We estimate a statistical model in a multimodal feature space for each possible query keyword. The key element of our framework is to identify feature space transformations that make them comparable in complexity and density. We select the optimal multimodal feature space with a minimum description length criterion from a set of candidate feature spaces that are computed with the average-mutual-information criterion for the text part and hierarchical expectation maximization for the visual part of the data. We evaluate our approach in three retrieval experiments (only text retrieval, only image retrieval, and text combined with image retrieval), verify the framework's low computational complexity, and compare with existing state-of-the-art ad-hoc models.

In large organizations the resources needed to solve challenging problems are typically dispersed over systems within and beyond the organization, and also in different media. However, there is still the need, in knowledge environments, for extraction methods able to combine evidence for a fact from across different media. In many cases the whole is more than the sum of its parts: only when considering the different media simultaneously can enough evidence be obtained to derive facts otherwise inaccessible to the knowledge worker via traditional methods that work on each single medium separately. In this paper, we present a cross-media knowledge extraction framework specifically designed to handle large volumes of documents composed of three types of media text, images and raw data and to exploit the evidence across the media. Our goal is to improve the quality and depth of automatically extracted knowledge.

In this paper we formulate image retrieval by text query as a vector space classification problem. This is achieved by creating a high-dimensional visual vocabulary that represents the image documents in great detail. We show how the representation of these image documents enables the application of well known text retrieval techniques such as Rocchio tf-idf and naíve Bayes to the semantic image retrieval problem. We tested these methods on a Corel images subset and achieve state-of-the-art retrieval performance using the proposed methods.

We propose a novel algorithm for extracting information by mining the feature space clusters and then assigning salient concepts to them. Bayesian techniques for extracting concepts from multimedia usually suffer either from lack of data or from too complex concepts to be represented by a single statistical model. An incremental information extraction approach, working at different levels of abstraction, would be able to handle concepts of varying complexities. We present the results of our research on the initial part of an incremental approach, the extraction of the most salient concepts from multimedia information.