Mariya Hendriksen

Mariya Hendriksen, Svitlana Vakulenko, Ernst Kuiper et al.

Most approaches to cross-modal retrieval (CMR) focus either on object-centric datasets, meaning that each document depicts or describes a single object, or on scene-centric datasets, meaning that each image depicts or describes a complex scene that involves multiple objects and relations between them. We posit that a robust CMR model should generalize well across both dataset types. Despite recent advances in CMR, the reproducibility of the results and their generalizability across different dataset types has not been studied before. We address this gap and focus on the reproducibility of the state-of-the-art CMR results when evaluated on object-centric and scene-centric datasets. We select two state-of-the-art CMR models with different architectures: (i) CLIP; and (ii) X-VLM. Additionally, we select two scene-centric datasets, and three object-centric datasets, and determine the relative performance of the selected models on these datasets. We focus on reproducibility, replicability, and generalizability of the outcomes of previously published CMR experiments. We discover that the experiments are not fully reproducible and replicable. Besides, the relative performance results partially generalize across object-centric and scene-centric datasets. On top of that, the scores obtained on object-centric datasets are much lower than the scores obtained on scene-centric datasets. For reproducibility and transparency we make our source code and the trained models publicly available.

Mariya Hendriksen, Shuo Zhang, Ridho Reinanda et al.

Image-Text Retrieval (ITR) systems are central to multimodal information access, with Vision-Language Models (VLMs) showing strong performance on standard benchmarks. However, these benchmarks predominantly rely on coarse-grained annotations, limiting their ability to reveal how models perform under real-world conditions, where query granularity varies. Motivated by this gap, we examine how dataset granularity and query perturbations affect retrieval performance and robustness across four architecturally diverse VLMs (ALIGN, AltCLIP, CLIP, and GroupViT). Using both standard benchmarks (MS-COCO, Flickr30k) and their fine-grained variants, we show that richer captions consistently enhance retrieval, especially in text-to-image tasks, where we observe an average improvement of 16.23%, compared to 6.44% in image-to-text. To assess robustness, we introduce a taxonomy of perturbations and conduct extensive experiments, revealing that while perturbations typically degrade performance, they can also unexpectedly improve retrieval, exposing nuanced model behaviors. Notably, word order emerges as a critical factor -- contradicting prior assumptions of model insensitivity to it. Our results highlight variation in model robustness and a dataset-dependent relationship between caption granularity and perturbation sensitivity and emphasize the necessity of evaluating models on datasets of varying granularity.

Thong Nguyen, Mariya Hendriksen, Andrew Yates et al.

Learned sparse retrieval (LSR) is a family of neural methods that encode queries and documents into sparse lexical vectors that can be indexed and retrieved efficiently with an inverted index. We explore the application of LSR to the multi-modal domain, with a focus on text-image retrieval. While LSR has seen success in text retrieval, its application in multimodal retrieval remains underexplored. Current approaches like LexLIP and STAIR require complex multi-step training on massive datasets. Our proposed approach efficiently transforms dense vectors from a frozen dense model into sparse lexical vectors. We address issues of high dimension co-activation and semantic deviation through a new training algorithm, using Bernoulli random variables to control query expansion. Experiments with two dense models (BLIP, ALBEF) and two datasets (MSCOCO, Flickr30k) show that our proposed algorithm effectively reduces co-activation and semantic deviation. Our best-performing sparsified model outperforms state-of-the-art text-image LSR models with a shorter training time and lower GPU memory requirements. Our approach offers an effective solution for training LSR retrieval models in multimodal settings. Our code and model checkpoints are available at github.com/thongnt99/lsr-multimodal

Mariya Hendriksen, Viggo Overes

The popularity of online fashion shopping continues to grow. The ability to offer an effective recommendation to customers is becoming increasingly important. In this work, we focus on Fashion Outfits Challenge, part of SIGIR 2022 Workshop on eCommerce. The challenge is centered around Fill in the Blank (FITB) task that implies predicting the missing outfit, given an incomplete outfit and a list of candidates. In this paper, we focus on applying siamese networks on the task. More specifically, we explore how combining information from multiple modalities (textual and visual modality) impacts the performance of the model on the task. We evaluate our model on the test split provided by the challenge organizers and the test split with gold assignments that we created during the development phase. We discover that using both visual, and visual and textual data demonstrates promising results on the task. We conclude by suggesting directions for further improvement of our method.

Kenneth Enevoldsen, Isaac Chung, Imene Kerboua et al. · cambridge, meta-ai

Text embeddings are typically evaluated on a limited set of tasks, which are constrained by language, domain, and task diversity. To address these limitations and provide a more comprehensive evaluation, we introduce the Massive Multilingual Text Embedding Benchmark (MMTEB) - a large-scale, community-driven expansion of MTEB, covering over 500 quality-controlled evaluation tasks across 250+ languages. MMTEB includes a diverse set of challenging, novel tasks such as instruction following, long-document retrieval, and code retrieval, representing the largest multilingual collection of evaluation tasks for embedding models to date. Using this collection, we develop several highly multilingual benchmarks, which we use to evaluate a representative set of models. We find that while large language models (LLMs) with billions of parameters can achieve state-of-the-art performance on certain language subsets and task categories, the best-performing publicly available model is multilingual-e5-large-instruct with only 560 million parameters. To facilitate accessibility and reduce computational cost, we introduce a novel downsampling method based on inter-task correlation, ensuring a diverse selection while preserving relative model rankings. Furthermore, we optimize tasks such as retrieval by sampling hard negatives, creating smaller but effective splits. These optimizations allow us to introduce benchmarks that drastically reduce computational demands. For instance, our newly introduced zero-shot English benchmark maintains a ranking order similar to the full-scale version but at a fraction of the computational cost.

Maurits Bleeker, Mariya Hendriksen, Andrew Yates et al.

Vision-language models (VLMs) mainly rely on contrastive training to learn general-purpose representations of images and captions. We focus on the situation when one image is associated with several captions, each caption containing both information shared among all captions and unique information per caption about the scene depicted in the image. In such cases, it is unclear whether contrastive losses are sufficient for learning task-optimal representations that contain all the information provided by the captions or whether the contrastive learning setup encourages the learning of a simple shortcut that minimizes contrastive loss. We introduce synthetic shortcuts for vision-language: a training and evaluation framework where we inject synthetic shortcuts into image-text data. We show that contrastive VLMs trained from scratch or fine-tuned with data containing these synthetic shortcuts mainly learn features that represent the shortcut. Hence, contrastive losses are not sufficient to learn task-optimal representations, i.e., representations that contain all task-relevant information shared between the image and associated captions. We examine two methods to reduce shortcut learning in our training and evaluation framework: (i) latent target decoding and (ii) implicit feature modification. We show empirically that both methods improve performance on the evaluation task, but only partly reduce shortcut learning when training and evaluating with our shortcut learning framework. Hence, we show the difficulty and challenge of our shortcut learning framework for contrastive vision-language representation learning.

Mariya Hendriksen, Svitlana Vakulenko, Jordan Massiah et al.

Large Language Models (LLMs) are transforming personalized search, recommendations, and customer interaction in e-commerce. Customers increasingly shop across multiple devices, from voice-only assistants to multimodal displays, each offering different input and output capabilities. A proactive suggestion to switch devices can greatly improve the user experience, but it must be offered with high precision to avoid unnecessary friction. We address the challenge of predicting when a query requires visual augmentation and a cross-device switch to improve product discovery. We introduce Image-Seeking Intent Prediction, a novel task for LLM-driven e-commerce assistants that anticipates when a spoken product query should proactively trigger a visual on a screen-enabled device. Using large-scale production data from a multi-device retail assistant, including 900K voice queries, associated product retrievals, and behavioral signals such as image carousel engagement, we train IRP (Image Request Predictor), a model that leverages user input query and corresponding retrieved product metadata to anticipate visual intent. Our experiments show that combining query semantics with product data, particularly when improved through lightweight summarization, consistently improves prediction accuracy. Incorporating a differentiable precision-oriented loss further reduces false positives. These results highlight the potential of LLMs to power intelligent, cross-device shopping assistants that anticipate and adapt to user needs, enabling more seamless and personalized e-commerce experiences.

Mariya Hendriksen, Tabish Rashid, David Bignell et al.

World models -- generative models that simulate environment dynamics conditioned on past observations and actions -- are gaining prominence in planning, simulation, and embodied AI. However, evaluating their rollouts remains a fundamental challenge, requiring fine-grained, temporally grounded assessment of action alignment and semantic consistency -- capabilities not captured by existing metrics. Vision-Language Models (VLMs) have shown promise as automatic evaluators of generative content due to their strong multimodal reasoning abilities. Yet, their use in fine-grained, temporally sensitive evaluation tasks remains limited and requires targeted adaptation. We introduce a evaluation protocol targeting two recognition tasks -- action recognition and character recognition -- each assessed across binary, multiple-choice, and open-ended formats. To support this, we present UNIVERSE (UNIfied Vision-language Evaluator for Rollouts in Simulated Environments), a method for adapting VLMs to rollout evaluation under data and compute constraints. We conduct a large-scale study comparing full, partial, and parameter-efficient finetuning across task formats, context lengths, sampling strategies, and data compositions. The resulting unified evaluator matches the performance of task-specific baselines using a single checkpoint. Human studies confirm strong alignment with human judgments, establishing UNIVERSE as a scalable, semantics-aware evaluator for world models.

Karel Veldkamp, Mariya Hendriksen, Zoltán Szlávik et al.

Contrastive learning is a powerful way of learning multimodal representations across various domains such as image-caption retrieval and audio-visual representation learning. In this work, we investigate if these findings generalize to the domain of music videos. Specifically, we create a dual en-coder for the audio and video modalities and train it using a bidirectional contrastive loss. For the experiments, we use an industry dataset containing 550 000 music videos as well as the public Million Song Dataset, and evaluate the quality of learned representations on the downstream tasks of music tagging and genre classification. Our results indicate that pre-trained networks without contrastive fine-tuning outperform our contrastive learning approach when evaluated on both tasks. To gain a better understanding of the reasons contrastive learning was not successful for music videos, we perform a qualitative analysis of the learned representations, revealing why contrastive learning might have difficulties uniting embeddings from two modalities. Based on these findings, we outline possible directions for future work. To facilitate the reproducibility of our results, we share our code and the pre-trained model.

Mariya Hendriksen, Maurits Bleeker, Svitlana Vakulenko et al.

E-commerce provides rich multimodal data that is barely leveraged in practice. One aspect of this data is a category tree that is being used in search and recommendation. However, in practice, during a user's session there is often a mismatch between a textual and a visual representation of a given category. Motivated by the problem, we introduce the task of category-to-image retrieval in e-commerce and propose a model for the task, CLIP-ITA. The model leverages information from multiple modalities (textual, visual, and attribute modality) to create product representations. We explore how adding information from multiple modalities (textual, visual, and attribute modality) impacts the model's performance. In particular, we observe that CLIP-ITA significantly outperforms a comparable model that leverages only the visual modality and a comparable model that leverages the visual and attribute modality.

Mariya Hendriksen, Ernst Kuiper, Pim Nauts et al.

The popularity of e-commerce platforms continues to grow. Being able to understand, and predict customer behavior is essential for customizing the user experience through personalized result presentations, recommendations, and special offers. Previous work has considered a broad range of prediction models as well as features inferred from clickstream data to record session characteristics, and features inferred from user data to record customer characteristics. So far, most previous work in the area of purchase prediction has focused on known customers, largely ignoring anonymous sessions, i.e., sessions initiated by a non-logged-in or unrecognized customer. However, in the de-identified data from a large European e-commerce platform available to us, more than 50% of the sessions start as anonymous sessions. In this paper, we focus on purchase prediction for both anonymous and identified sessions on an e-commerce platform. We start with a descriptive analysis of purchase vs. non-purchase sessions. This analysis informs the definition of a feature-based model for purchase prediction for anonymous sessions and identified sessions; our models consider a range of session-based features for anonymous sessions, such as the channel type, the number of visited pages, and the device type. For identified user sessions, our analysis points to customer history data as a valuable discriminator between purchase and non-purchase sessions. Based on our analysis, we build two types of predictors: (1) a predictor for anonymous that beats a production-ready predictor by over 17.54% F1; and (2) a predictor for identified customers that uses session data as well as customer history and achieves an F1 of 96.20%. Finally, we discuss the broader practical implications of our findings.