Markus Schedl

Deepak Kumar, Oleg Lesota, George Zerveas et al. · microsoft-research

Large pre-trained language models contain societal biases and carry along these biases to downstream tasks. Current in-processing bias mitigation approaches (like adversarial training) impose debiasing by updating a model's parameters, effectively transferring the model to a new, irreversible debiased state. In this work, we propose a novel approach to develop stand-alone debiasing functionalities separate from the model, which can be integrated into the model on-demand, while keeping the core model untouched. Drawing from the concept of AdapterFusion in multi-task learning, we introduce DAM (Debiasing with Adapter Modules) - a debiasing approach to first encapsulate arbitrary bias mitigation functionalities into separate adapters, and then add them to the model on-demand in order to deliver fairness qualities. We conduct a large set of experiments on three classification tasks with gender, race, and age as protected attributes. Our results show that DAM improves or maintains the effectiveness of bias mitigation, avoids catastrophic forgetting in a multi-attribute scenario, and maintains on-par task performance, while granting parameter-efficiency and easy switching between the original and debiased models.

Christian Ganhör, David Penz, Navid Rekabsaz et al.

Collaborative filtering algorithms capture underlying consumption patterns, including the ones specific to particular demographics or protected information of users, e.g. gender, race, and location. These encoded biases can influence the decision of a recommendation system (RS) towards further separation of the contents provided to various demographic subgroups, and raise privacy concerns regarding the disclosure of users' protected attributes. In this work, we investigate the possibility and challenges of removing specific protected information of users from the learned interaction representations of a RS algorithm, while maintaining its effectiveness. Specifically, we incorporate adversarial training into the state-of-the-art MultVAE architecture, resulting in a novel model, Adversarial Variational Auto-Encoder with Multinomial Likelihood (Adv-MultVAE), which aims at removing the implicit information of protected attributes while preserving recommendation performance. We conduct experiments on the MovieLens-1M and LFM-2b-DemoBias datasets, and evaluate the effectiveness of the bias mitigation method based on the inability of external attackers in revealing the users' gender information from the model. Comparing with baseline MultVAE, the results show that Adv-MultVAE, with marginal deterioration in performance (w.r.t. NDCG and recall), largely mitigates inherent biases in the model on both datasets.

Lukas Hauzenberger, Shahed Masoudian, Deepak Kumar et al.

Societal biases are reflected in large pre-trained language models and their fine-tuned versions on downstream tasks. Common in-processing bias mitigation approaches, such as adversarial training and mutual information removal, introduce additional optimization criteria, and update the model to reach a new debiased state. However, in practice, end-users and practitioners might prefer to switch back to the original model, or apply debiasing only on a specific subset of protected attributes. To enable this, we propose a novel modular bias mitigation approach, consisting of stand-alone highly sparse debiasing subnetworks, where each debiasing module can be integrated into the core model on-demand at inference time. Our approach draws from the concept of \emph{diff} pruning, and proposes a novel training regime adaptable to various representation disentanglement optimizations. We conduct experiments on three classification tasks with gender, race, and age as protected attributes. The results show that our modular approach, while maintaining task performance, improves (or at least remains on-par with) the effectiveness of bias mitigation in comparison with baseline finetuning. Particularly on a two-attribute dataset, our approach with separately learned debiasing subnetworks shows effective utilization of either or both the subnetworks for selective bias mitigation.

Dominik Kowald, Gregor Mayr, Markus Schedl et al.

Recent research has suggested different metrics to measure the inconsistency of recommendation performance, including the accuracy difference between user groups, miscalibration, and popularity lift. However, a study that relates miscalibration and popularity lift to recommendation accuracy across different user groups is still missing. Additionally, it is unclear if particular genres contribute to the emergence of inconsistency in recommendation performance across user groups. In this paper, we present an analysis of these three aspects of five well-known recommendation algorithms for user groups that differ in their preference for popular content. Additionally, we study how different genres affect the inconsistency of recommendation performance, and how this is aligned with the popularity of the genres. Using data from LastFm, MovieLens, and MyAnimeList, we present two key findings. First, we find that users with little interest in popular content receive the worst recommendation accuracy, and that this is aligned with miscalibration and popularity lift. Second, our experiments show that particular genres contribute to a different extent to the inconsistency of recommendation performance, especially in terms of miscalibration in the case of the MyAnimeList dataset.

Passant Elchafei, Monorama Swain, Shahed Masoudian et al.

Retrieval-Augmented Generation (RAG) aims to reduce hallucination by grounding answers in retrieved evidence, yet hallucinated answers remain common even when relevant documents are available. Existing evaluations focus on answer-level or passage-level accuracy, offering limited insight into how evidence is used during generation. In this work, we introduce a facet-level diagnostics framework for QA that decomposes each input question into atomic reasoning facets. For each facet, we assess evidence sufficiency and grounding using a structured Facet x Chunk matrix that combines retrieval relevance with natural language inference-based faithfulness scores. To diagnose evidence usage, we analyze three controlled inference modes: Strict RAG, which enforces exclusive reliance on retrieved evidence; Soft RAG, which allows integration of retrieved evidence and parametric knowledge; and LLM-only generation without retrieval. Comparing these modes enables thorough analysis of retrieval-generation misalignment, defined as cases where relevant evidence is retrieved but not correctly integrated during generation. Across medical QA and HotpotQA, we evaluate three open-source and closed-source LLMs (GPT, Gemini, and LLaMA), providing interpretable diagnostics that reveal recurring facet-level failure modes, including evidence absence, evidence misalignment, and prior-driven overrides. Our results demonstrate that hallucinations in RAG systems are driven less by retrieval accuracy and more by how retrieved evidence is integrated during generation, with facet-level analysis exposing systematic evidence override and misalignment patterns that remain hidden under answer-level evaluation.

Shahed Masoudian, Khaled Koutini, Markus Schedl et al.

Domain shift is considered a challenge in machine learning as it causes significant degradation of model performance. In the Acoustic Scene Classification task (ASC), domain shift is mainly caused by different recording devices. Several studies have already targeted domain generalization to improve the performance of ASC models on unseen domains, such as new devices. Recently, the Controllable Gate Adapter ConGater has been proposed in Natural Language Processing to address the biased training data problem. ConGater allows controlling the debiasing process at inference time. ConGater's main advantage is the continuous and selective debiasing of a trained model, during inference. In this work, we adapt ConGater to the audio spectrogram transformer for an acoustic scene classification task. We show that ConGater can be used to selectively adapt the learned representations to be invariant to device domain shifts such as recording devices. Our analysis shows that ConGater can progressively remove device information from the learned representations and improve the model generalization, especially under domain shift conditions (e.g. unseen devices). We show that information removal can be extended to both device and location domain. Finally, we demonstrate ConGater's ability to enhance specific device performance without further training.

Deepak Kumar, Emmanouil Karystinaios, Gerhard Widmer et al.

Music often shares notable parallels with language, motivating the use of pretrained large language models (LLMs) for symbolic music understanding and generation. Despite growing interest, the practical effectiveness of adapting instruction-tuned LLMs to symbolic music remains insufficiently characterized. We present a controlled comparative study of finetuning strategies for ABC-based generation and understanding, comparing an off-the-shelf instruction-tuned backbone to domain-adapted variants and a music-specialized LLM baseline. Across multiple symbolic music corpora and evaluation signals, we provide some insights into adaptation choices for symbolic music applications. We highlight the domain adaptation vs.~preserving prior information tradeoff as well as the distinct behaviour of metrics used to measure the domain adaptation for symbolic music.

Shahed Masoudian, Markus Frohmann, Navid Rekabsaz et al.

Language models frequently inherit societal biases from their training data. Numerous techniques have been proposed to mitigate these biases during both the pre-training and fine-tuning stages. However, fine-tuning a pre-trained debiased language model on a downstream task can reintroduce biases into the model. Additionally, existing debiasing methods for downstream tasks either (i) require labels of protected attributes (e.g., age, race, or political views) that are often not available or (ii) rely on indicators of bias, which restricts their applicability to gender debiasing since they rely on gender-specific words. To address this, we introduce a novel debiasing regularization technique based on the class-wise variance of embeddings. Crucially, our method does not require attribute labels and targets any attribute, thus addressing the shortcomings of existing debiasing methods. Our experiments on encoder language models and three datasets demonstrate that our method outperforms existing strong debiasing baselines that rely on target attribute labels while maintaining performance on the target task.

Christian Ganhör, Marta Moscati, Anna Hausberger et al.

Most recommender systems adopt collaborative filtering (CF) and provide recommendations based on past collective interactions. Therefore, the performance of CF algorithms degrades when few or no interactions are available, a scenario referred to as cold-start. To address this issue, previous work relies on models leveraging both collaborative data and side information on the users or items. Similar to multimodal learning, these models aim at combining collaborative and content representations in a shared embedding space. In this work we propose a novel technique for multimodal recommendation, relying on a multimodal Single-Branch embedding network for Recommendation (SiBraR). Leveraging weight-sharing, SiBraR encodes interaction data as well as multimodal side information using the same single-branch embedding network on different modalities. This makes SiBraR effective in scenarios of missing modality, including cold start. Our extensive experiments on large-scale recommendation datasets from three different recommendation domains (music, movie, and e-commerce) and providing multimodal content information (audio, text, image, labels, and interactions) show that SiBraR significantly outperforms CF as well as state-of-the-art content-based RSs in cold-start scenarios, and is competitive in warm scenarios. We show that SiBraR's recommendations are accurate in missing modality scenarios, and that the model is able to map different modalities to the same region of the shared embedding space, hence reducing the modality gap.

Zihan Li, Gustavo Escobedo, Marta Moscati et al.

Diffusion recommender systems achieve strong recommendation accuracy but often suffer from popularity bias, resulting in unequal item exposure. To address this shortcoming, we introduce A2G-DiffRec, a diffusion recommender that incorporates adaptive autoguidance, where the main model is guided by a less-trained version of itself. Instead of using a fixed guidance weight, A2G-DiffRec learns to adaptively weigh the outputs of the main and weak models during training, supervised by a fairness-aware regularization that promotes balanced exposure across items with different popularity levels. Experimental results on three public datasets show that A2G-DiffRec is effective in enhancing item-side fairness at a marginal cost of accuracy reduction compared to existing guided diffusion recommenders and other non-diffusion baselines.

Abdul Hannan, Daniele Falavigna, Shah Nawaz et al.

Edge devices operate in constrained and varying resource settings, requiring dynamic architectures that can adapt to limitations of the available resources. To meet such demands, layer dropping ($\mathcal{LD}$) approach is typically used to transform static models into dynamic ones by skipping parts of the network along with reducing overall computational complexity. However, existing $\mathcal{LD}$ methods greatly impact the dynamic model's performance for low and high dropping cases, deteriorating the performance-computation trade-off. To this end, we propose a distillation-based layer dropping (DLD) framework that effectively combines the capabilities of knowledge distillation and $\mathcal{LD}$ in an end-to-end fashion, thereby achieving state-of-the-art performance for dynamic speech networks. Comprehensive experimentation utilizing well-known speech recognition methods, including conformer and WavLM, on three public benchmarks demonstrates the effectiveness of our framework, reducing the word error rate by $9.32\%$ and $2.25\%$ for high and no dropping cases with $33.3\%$ reduction in training time.

Muhammad Saad Saeed, Shah Nawaz, Muhammad Zaigham Zaheer et al.

Multimodal networks have demonstrated remarkable performance improvements over their unimodal counterparts. Existing multimodal networks are designed in a multi-branch fashion that, due to the reliance on fusion strategies, exhibit deteriorated performance if one or more modalities are missing. In this work, we propose a modality invariant multimodal learning method, which is less susceptible to the impact of missing modalities. It consists of a single-branch network sharing weights across multiple modalities to learn inter-modality representations to maximize performance as well as robustness to missing modalities. Extensive experiments are performed on four challenging datasets including textual-visual (UPMC Food-101, Hateful Memes, Ferramenta) and audio-visual modalities (VoxCeleb1). Our proposed method achieves superior performance when all modalities are present as well as in the case of missing modalities during training or testing compared to the existing state-of-the-art methods.

Marta Moscati, Muhammad Saad Saeed, Marina Zanoni et al.

Multimodal speaker identification systems typically assume the availability of complete and homogeneous audio-visual modalities during both training and testing. However, in real-world applications, such assumptions often do not hold. Visual information may be missing due to occlusions, camera failures, or privacy constraints, while multilingual speakers introduce additional complexity due to linguistic variability across languages. These challenges significantly affect the robustness and generalization of multimodal speaker identification systems. The POLY-SIM Grand Challenge 2026 aims to advance research in multimodal speaker identification under missing-modality and cross-lingual conditions. Specifically, the Grand Challenge encourages the development of robust methods that can effectively leverage incomplete multimodal inputs while maintaining strong performance across different languages. This report presents the design and organization of the POLY-SIM Grand Challenge 2026, including the dataset, task formulation, evaluation protocol, and baseline model. By providing a standardized benchmark and evaluation framework, the challenge aims to foster progress toward more robust and practical multimodal speaker identification systems.

Marta Moscati, Oleksandr Kats, Mubashir Noman et al.

Face-voice association is widely studied in multimodal learning and is approached representing faces and voices with embeddings that are close for a same person and well separated from those of others. Previous work achieved this with loss functions. Recent advancements in classification have shown that the discriminative ability of embeddings can be strengthened by imposing maximum class separation as inductive bias. This technique has never been used in the domain of face-voice association, and this work aims at filling this gap. More specifically, we develop a method for face-voice association that imposes maximum class separation among multimodal representations of different speakers as an inductive bias. Through quantitative experiments we demonstrate the effectiveness of our approach, showing that it achieves SOTA performance on two task formulation of face-voice association. Furthermore, we carry out an ablation study to show that imposing inductive bias is most effective when combined with losses for inter-class orthogonality. To the best of our knowledge, this work is the first that applies and demonstrates the effectiveness of maximum class separation as an inductive bias in multimodal learning; it hence paves the way to establish a new paradigm.

Marta Moscati, Ahmed Abdullah, Muhammad Saad Saeed et al.

Over half of the world's population is bilingual and people often communicate under multilingual scenarios. The Face-Voice Association in Multilingual Environments (FAME) 2026 Challenge, held at ICASSP 2026, focuses on developing methods for face-voice association that are effective when the language at test-time is different than the training one. This report provides a brief summary of the challenge.

Salem AlMarri, Muhammad Irzam Liaqat, Muhammad Zaigham Zaheer et al.

In recent years, multimodal anomaly detection methods have demonstrated remarkable performance improvements over video-only models. However, real-world multimodal data is often corrupted due to unforeseen environmental distortions. In this paper, we present the first-of-its-kind work that comprehensively investigates the adverse effects of corrupted modalities on multimodal anomaly detection task. To streamline this work, we propose RobustA, a carefully curated evaluation dataset to systematically observe the impacts of audio and visual corruptions on the overall effectiveness of anomaly detection systems. Furthermore, we propose a multimodal anomaly detection method, which shows notable resilience against corrupted modalities. The proposed method learns a shared representation space for different modalities and employs a dynamic weighting scheme during inference based on the estimated level of corruption. Our work represents a significant step forward in enabling the real-world application of multimodal anomaly detection, addressing situations where the likely events of modality corruptions occur. The proposed evaluation dataset with corrupted modalities and respective extracted features will be made publicly available.

Markus Essl, Marta Moscati, Mubashir Noman et al.

Multimodal sensor fusion has demonstrated remarkable performance improvements over unimodal approaches in 3D object detection for autonomous vehicles. Typically, existing methods transform multimodal data from independent sensors, such as camera and LiDAR, into a unified bird's-eye view (BEV) representation for fusion. Although effective in ideal conditions, this strategy suffers from substantial performance deterioration when camera or LiDAR data are missing, corrupted, or noisy. To address this vulnerability, we develop a framework-agnostic fusion module for camera and LiDAR data that allows for handling cases when one of the two modalities is missing or corrupted. To demonstrate the effectiveness of our module, we instantiate it in BEVFusion [1], a well-established framework to combine camera and LiDAR data for 3D object detection. By means of quantitative experiments on the MultiCorrupt dataset, we demonstrate that our module achieves favorable performance improvements under scenarios of missing and corrupted modalities, substantially outperforming existing unified representation approaches across a wide range of sensor deterioration scenarios and reaching state-of-the-art performance in scenarios of corrupted modality due to extreme weather conditions and sensor failure.

Markus Frohmann, Elena V. Epure, Gabriel Meseguer-Brocal et al.

The recent rise in capabilities of AI-based music generation tools has created an upheaval in the music industry, necessitating the creation of accurate methods to detect such AI-generated content. This can be done using audio-based detectors; however, it has been shown that they struggle to generalize to unseen generators or when the audio is perturbed. Furthermore, recent work used accurate and cleanly formatted lyrics sourced from a lyrics provider database to detect AI-generated music. However, in practice, such perfect lyrics are not available (only the audio is); this leaves a substantial gap in applicability in real-life use cases. In this work, we instead propose solving this gap by transcribing songs using general automatic speech recognition (ASR) models. We do this using several detectors. The results on diverse, multi-genre, and multi-lingual lyrics show generally strong detection performance across languages and genres, particularly for our best-performing model using Whisper large-v2 and LLM2Vec embeddings. In addition, we show that our method is more robust than state-of-the-art audio-based ones when the audio is perturbed in different ways and when evaluated on different music generators. Our code is available at https://github.com/deezer/robust-AI-lyrics-detection.

Markus Frohmann, Gabriel Meseguer-Brocal, Markus Schedl et al.

The rapid advancement of AI-based music generation tools is revolutionizing the music industry but also posing challenges to artists, copyright holders, and providers alike. This necessitates reliable methods for detecting such AI-generated content. However, existing detectors, relying on either audio or lyrics, face key practical limitations: audio-based detectors fail to generalize to new or unseen generators and are vulnerable to audio perturbations; lyrics-based methods require cleanly formatted and accurate lyrics, unavailable in practice. To overcome these limitations, we propose a novel, practically grounded approach: a multimodal, modular late-fusion pipeline that combines automatically transcribed sung lyrics and speech features capturing lyrics-related information within the audio. By relying on lyrical aspects directly from audio, our method enhances robustness, mitigates susceptibility to low-level artifacts, and enables practical applicability. Experiments show that our method, DE-detect, outperforms existing lyrics-based detectors while also being more robust to audio perturbations. Thus, it offers an effective, robust solution for detecting AI-generated music in real-world scenarios. Our code is available at https://github.com/deezer/robust-AI-lyrics-detection.

Markus Frohmann, Igor Sterner, Ivan Vulić et al.

Segmenting text into sentences plays an early and crucial role in many NLP systems. This is commonly achieved by using rule-based or statistical methods relying on lexical features such as punctuation. Although some recent works no longer exclusively rely on punctuation, we find that no prior method achieves all of (i) robustness to missing punctuation, (ii) effective adaptability to new domains, and (iii) high efficiency. We introduce a new model - Segment any Text (SaT) - to solve this problem. To enhance robustness, we propose a new pretraining scheme that ensures less reliance on punctuation. To address adaptability, we introduce an extra stage of parameter-efficient fine-tuning, establishing state-of-the-art performance in distinct domains such as verses from lyrics and legal documents. Along the way, we introduce architectural modifications that result in a threefold gain in speed over the previous state of the art and solve spurious reliance on context far in the future. Finally, we introduce a variant of our model with fine-tuning on a diverse, multilingual mixture of sentence-segmented data, acting as a drop-in replacement and enhancement for existing segmentation tools. Overall, our contributions provide a universal approach for segmenting any text. Our method outperforms all baselines - including strong LLMs - across 8 corpora spanning diverse domains and languages, especially in practically relevant situations where text is poorly formatted. Our models and code, including documentation, are available at https://github.com/segment-any-text/wtpsplit under the MIT license.

Klara Krieg, Emilia Parada-Cabaleiro, Gertraud Medicus et al.

The provided contents by information retrieval (IR) systems can reflect the existing societal biases and stereotypes. Such biases in retrieval results can lead to further establishing and strengthening stereotypes in society and also in the systems. To facilitate the studies of gender bias in the retrieval results of IR systems, we introduce Gender Representation-Bias for Information Retrieval (Grep-BiasIR), a novel thoroughly-audited dataset consisting of 118 bias-sensitive neutral search queries. The set of queries covers a wide range of gender-related topics, for which a biased representation of genders in the search result can be considered as socially problematic. Each query is accompanied with one relevant and one non-relevant document, where the document is also provided in three variations of female, male, and neutral. The dataset is available at https://github.com/KlaraKrieg/GrepBiasIR.

Passant Elchafei, Hossam Emam, Mohamed Alansary et al.

We present H-RAG, our submission to SemEval-2026 Task 8 (MTRAGEval), addressing both Task A (Retrieval) and Task C (Generation with Retrieved Passages). Task A evaluates standalone retrieval quality, while Task C assesses end-to-end retrieval-augmented generation (RAG) in multi-turn conversational settings, requiring both accurate answer generation and faithful grounding in retrieved evidence. Our approach implements a hierarchical parent-child RAG pipeline that separates fine-grained child-level retrieval from parent-level context reconstruction during generation. Documents are segmented into overlapping sentence-based child chunks, while full documents are preserved as parent units to provide coherent context. Retrieval combines hybrid dense-sparse search, tunable weighting, and embedding-based similarity rescoring over child chunks. Retrieved evidence is aggregated at the parent level and supplied to an instruction-tuned language model for response generation. H-RAG achieves an nDCG@5 score of 0.4271 on Task A and a harmonic mean score of 0.3241 on Task C (RB_agg: 0.2488, RL_F: 0.2703, RB_llm: 0.6508), underscoring the importance of retrieval configuration and parent-level aggregation in multi-turn RAG performance.

Peter Müllner, Dominik Kowald, Markus Schedl et al.

Balancing differential privacy (DP) with recommendation accuracy is a key challenge in privacy-preserving recommender systems, since DP-noise degrades accuracy. We address this trade-off at both the data and model levels. At the data level, we apply DP only to the most stereotypical user data likely to reveal sensitive attributes, such as gender or age, to reduce unnecessary perturbation; we refer to this as targeted DP. At the model level, we use meta-learning to improve robustness to remaining DP-noise. This achieves a better trade-off between accuracy and privacy than standard approaches: Meta-learning improves accuracy and targeted DP leads to lower empirical privacy risk compared to uniformly applied DP and full DP baselines. Overall, our findings show that selectively applying DP at the data level together with meta-learning at the model level can effectively balance recommendation accuracy and user privacy.

Shahed Masoudian, Cornelia Volaucnik, Markus Schedl et al.

Bias mitigation of Language Models has been the topic of many studies with a recent focus on learning separate modules like adapters for on-demand debiasing. Besides optimizing for a modularized debiased model, it is often critical in practice to control the degree of bias reduction at inference time, e.g., in order to tune for a desired performance-fairness trade-off in search results or to control the strength of debiasing in classification tasks. In this paper, we introduce Controllable Gate Adapter (ConGater), a novel modular gating mechanism with adjustable sensitivity parameters, which allows for a gradual transition from the biased state of the model to the fully debiased version at inference time. We demonstrate ConGater performance by (1) conducting adversarial debiasing experiments with three different models on three classification tasks with four protected attributes, and (2) reducing the bias of search results through fairness list-wise regularization to enable adjusting a trade-off between performance and fairness metrics. Our experiments on the classification tasks show that compared to baselines of the same caliber, ConGater can maintain higher task performance while containing less information regarding the attributes. Our results on the retrieval task show that the fully debiased ConGater can achieve the same fairness performance while maintaining more than twice as high task performance than recent strong baselines. Overall, besides strong performance ConGater enables the continuous transitioning between biased and debiased states of models, enhancing personalization of use and interpretability through controllability.

Gustavo Escobedo, Christian Ganhör, Stefan Brandl et al.

In widely used neural network-based collaborative filtering models, users' history logs are encoded into latent embeddings that represent the users' preferences. In this setting, the models are capable of mapping users' protected attributes (e.g., gender or ethnicity) from these user embeddings even without explicit access to them, resulting in models that may treat specific demographic user groups unfairly and raise privacy issues. While prior work has approached the removal of a single protected attribute of a user at a time, multiple attributes might come into play in real-world scenarios. In the work at hand, we present AdvXMultVAE which aims to unlearn multiple protected attributes (exemplified by gender and age) simultaneously to improve fairness across demographic user groups. For this purpose, we couple a variational autoencoder (VAE) architecture with adversarial training (AdvMultVAE) to support simultaneous removal of the users' protected attributes with continuous and/or categorical values. Our experiments on two datasets, LFM-2b-100k and Ml-1m, from the music and movie domains, respectively, show that our approach can yield better results than its singular removal counterparts (based on AdvMultVAE) in effectively mitigating demographic biases whilst improving the anonymity of latent embeddings.

Muhammad Saad Saeed, Shah Nawaz, Muhammad Salman Tahir et al.

The advancements of technology have led to the use of multimodal systems in various real-world applications. Among them, the audio-visual systems are one of the widely used multimodal systems. In the recent years, associating face and voice of a person has gained attention due to presence of unique correlation between them. The Face-voice Association in Multilingual Environments (FAME) Challenge 2024 focuses on exploring face-voice association under a unique condition of multilingual scenario. This condition is inspired from the fact that half of the world's population is bilingual and most often people communicate under multilingual scenario. The challenge uses a dataset namely, Multilingual Audio-Visual (MAV-Celeb) for exploring face-voice association in multilingual environments. This report provides the details of the challenge, dataset, baselines and task details for the FAME Challenge.

Abdul Hannan, Muhammad Arslan Manzoor, Shah Nawaz et al.

We study the task of learning association between faces and voices, which is gaining interest in the multimodal community lately. These methods suffer from the deliberate crafting of negative mining procedures as well as the reliance on the distant margin parameter. These issues are addressed by learning a joint embedding space in which orthogonality constraints are applied to the fused embeddings of faces and voices. However, embedding spaces of faces and voices possess different characteristics and require spaces to be aligned before fusing them. To this end, we propose a method that accurately aligns the embedding spaces and fuses them with an enhanced gated fusion thereby improving the performance of face-voice association. Extensive experiments on the VoxCeleb dataset reveals the merits of the proposed approach.

Alessandro B. Melchiorre, Elena V. Epure, Shahed Masoudian et al.

Natural language interfaces offer a compelling approach for music recommendation, enabling users to express complex preferences conversationally. While Large Language Models (LLMs) show promise in this direction, their scalability in recommender systems is limited by high costs and latency. Retrieval-based approaches using smaller language models mitigate these issues but often rely on single-modal item representations, overlook long-term user preferences, and require full model retraining, posing challenges for real-world deployment. In this paper, we present JAM (Just Ask for Music), a lightweight and intuitive framework for natural language music recommendation. JAM models user-query-item interactions as vector translations in a shared latent space, inspired by knowledge graph embedding methods like TransE. To capture the complexity of music and user intent, JAM aggregates multimodal item features via cross-attention and sparse mixture-of-experts. We also introduce JAMSessions, a new dataset of over 100k user-query-item triples with anonymized user/item embeddings, uniquely combining conversational queries and user long-term preferences. Our results show that JAM provides accurate recommendations, produces intuitive representations suitable for practical use cases, and can be easily integrated with existing music recommendation stacks.

Felix Breiteneder, Mohammad Belal, Muhammad Saad Saeed et al.

Internet memes are powerful tools for communication, capable of spreading political, psychological, and sociocultural ideas. However, they can be harmful and can be used to disseminate hate toward targeted individuals or groups. Although previous studies have focused on designing new detection methods, these often rely on modal-complete data, such as text and images. In real-world settings, however, modalities like text may be missing due to issues like poor OCR quality, making existing methods sensitive to missing information and leading to performance deterioration. To address this gap, in this paper, we present the first-of-its-kind work to comprehensively investigate the behavior of harmful meme detection methods in the presence of modal-incomplete data. Specifically, we propose a new baseline method that learns a shared representation for multiple modalities by projecting them independently. These shared representations can then be leveraged when data is modal-incomplete. Experimental results on two benchmark datasets demonstrate that our method outperforms existing approaches when text is missing. Moreover, these results suggest that our method allows for better integration of visual features, reducing dependence on text and improving robustness in scenarios where textual information is missing. Our work represents a significant step forward in enabling the real-world application of harmful meme detection, particularly in situations where a modality is absent.

Elena V. Epure, Yashar Deldjoo, Bruno Sguerra et al.

Music Recommender Systems (MRS) have long relied on an information-retrieval framing, where progress is measured mainly through accuracy on retrieval-oriented subtasks. While effective, this reductionist paradigm struggles to address the deeper question of what makes a good recommendation, and attempts to broaden evaluation, through user studies or fairness analyses, have had limited impact. The emergence of Large Language Models (LLMs) disrupts this framework: LLMs are generative rather than ranking-based, making standard accuracy metrics questionable. They also introduce challenges such as hallucinations, knowledge cutoffs, non-determinism, and opaque training data, rendering traditional train/test protocols difficult to interpret. At the same time, LLMs create new opportunities, enabling natural-language interaction and even allowing models to act as evaluators. This work argues that the shift toward LLM-driven MRS requires rethinking evaluation. We first review how LLMs reshape user modeling, item modeling, and natural-language recommendation in music. We then examine evaluation practices from NLP, highlighting methodologies and open challenges relevant to MRS. Finally, we synthesize insights-focusing on how LLM prompting applies to MRS, to outline a structured set of success and risk dimensions. Our goal is to provide the MRS community with an updated, pedagogical, and cross-disciplinary perspective on evaluation.

Monorama Swain, Bubai Maji, Jagabandhu Mishra et al.

In this work, we address the challenge of building fair English ASR systems for second-language speakers. Our analysis of widely used ASR models, Whisper and Seamless-M4T, reveals large fluctuations in word error rate (WER) across 26 accent groups, indicating significant fairness gaps. To mitigate this, we propose fairness-prompted finetuning with lightweight adapters, incorporating Spectral Decoupling (SD), Group Distributionally Robust Optimization (Group-DRO), and Invariant Risk Minimization (IRM). Our proposed fusion of traditional empirical risk minimization (ERM) with cross-entropy and fairness-driven objectives (SD, Group DRO, and IRM) enhances fairness across accent groups while maintaining overall recognition accuracy. In terms of macro-averaged word error rate, our approach achieves a relative improvement of 58.7% and 58.5% over the large pretrained Whisper and SeamlessM4T, and 9.7% and 7.8% over them, finetuning with standard empirical risk minimization with cross-entropy loss.

Marta Moscati, Ahmed Abdullah, Muhammad Saad Saeed et al.

The advancements of technology have led to the use of multimodal systems in various real-world applications. Among them, audio-visual systems are among the most widely used multimodal systems. In the recent years, associating face and voice of a person has gained attention due to the presence of unique correlation between them. The Face-voice Association in Multilingual Environments (FAME) 2026 Challenge focuses on exploring face-voice association under the unique condition of a multilingual scenario. This condition is inspired from the fact that half of the world's population is bilingual and most often people communicate under multilingual scenarios. The challenge uses a dataset named Multilingual Audio-Visual (MAV-Celeb) for exploring face-voice association in multilingual environments. This report provides the details of the challenge, dataset, baseline models, and task details for the FAME Challenge.

Marta Moscati, Shah Nawaz, Markus Schedl

Recommender Systems (RS) often rely on representations of users and items in a joint embedding space and on a similarity metric to compute relevance scores. In modern RS, the modules to obtain user and item representations consist of two distinct and separate neural networks (NN). In multimodal representation learning, weight sharing has been proven effective in reducing the distance between multiple modalities of a same item. Inspired by these approaches, we propose a novel RS that leverages weight sharing between the user and item NN modules used to obtain the latent representations in the shared embedding space. The proposed framework consists of a single Collaborative Branch for Recommendation (CoBraR). We evaluate CoBraR by means of quantitative experiments on e-commerce and movie recommendation. Our experiments show that by reducing the number of parameters and improving beyond-accuracy aspects without compromising accuracy, CoBraR has the potential to be applied and extended for real-world scenarios.

Shahed Masoudian, Gustavo Escobedo, Hannah Strauss et al.

As Large Language Models (LLMs) are increasingly used across different applications, concerns about their potential to amplify gender biases in various tasks are rising. Prior research has often probed gender bias using explicit gender cues as counterfactual, or studied them in sentence completion and short question answering tasks. These formats might overlook more implicit forms of bias embedded in generative behavior of longer content. In this work, we investigate gender bias in LLMs using gender stereotypes studied in psychology (e.g., aggressiveness or gossiping) in an open-ended task of narrative generation. We introduce a novel dataset called StereoBias-Stories containing short stories either unconditioned or conditioned on (one, two, or six) random attributes from 25 psychological stereotypes and three task-related story endings. We analyze how the gender contribution in the overall story changes in response to these attributes and present three key findings: (1) While models, on average, are highly biased towards male in unconditioned prompts, conditioning on attributes independent from gender stereotypes mitigates this bias. (2) Combining multiple attributes associated with the same gender stereotype intensifies model behavior, with male ones amplifying bias and female ones alleviating it. (3) Model biases align with psychological ground-truth used for categorization, and alignment strength increases with model size. Together, these insights highlight the importance of psychology-grounded evaluation of LLMs.

Darius Afchar, Alessandro B. Melchiorre, Markus Schedl et al.

The most common way to listen to recorded music nowadays is via streaming platforms which provide access to tens of millions of tracks. To assist users in effectively browsing these large catalogs, the integration of Music Recommender Systems (MRSs) has become essential. Current real-world MRSs are often quite complex and optimized for recommendation accuracy. They combine several building blocks based on collaborative filtering and content-based recommendation. This complexity can hinder the ability to explain recommendations to end users, which is particularly important for recommendations perceived as unexpected or inappropriate. While pure recommendation performance often correlates with user satisfaction, explainability has a positive impact on other factors such as trust and forgiveness, which are ultimately essential to maintain user loyalty. In this article, we discuss how explainability can be addressed in the context of MRSs. We provide perspectives on how explainability could improve music recommendation algorithms and enhance user experience. First, we review common dimensions and goals of recommenders' explainability and in general of eXplainable Artificial Intelligence (XAI), and elaborate on the extent to which these apply -- or need to be adapted -- to the specific characteristics of music consumption and recommendation. Then, we show how explainability components can be integrated within a MRS and in what form explanations can be provided. Since the evaluation of explanation quality is decoupled from pure accuracy-based evaluation criteria, we also discuss requirements and strategies for evaluating explanations of music recommendations. Finally, we describe the current challenges for introducing explainability within a large-scale industrial music recommender system and provide research perspectives.

Oleg Lesota, Alessandro B. Melchiorre, Navid Rekabsaz et al.

Several studies have identified discrepancies between the popularity of items in user profiles and the corresponding recommendation lists. Such behavior, which concerns a variety of recommendation algorithms, is referred to as popularity bias. Existing work predominantly adopts simple statistical measures, such as the difference of mean or median popularity, to quantify popularity bias. Moreover, it does so irrespective of user characteristics other than the inclination to popular content. In this work, in contrast, we propose to investigate popularity differences (between the user profile and recommendation list) in terms of median, a variety of statistical moments, as well as similarity measures that consider the entire popularity distributions (Kullback-Leibler divergence and Kendall's tau rank-order correlation). This results in a more detailed picture of the characteristics of popularity bias. Furthermore, we investigate whether such algorithmic popularity bias affects users of different genders in the same way. We focus on music recommendation and conduct experiments on the recently released standardized LFM-2b dataset, containing listening profiles of Last.fm users. We investigate the algorithmic popularity bias of seven common recommendation algorithms (five collaborative filtering and two baselines). Our experiments show that (1) the studied metrics provide novel insights into popularity bias in comparison with only using average differences, (2) algorithms less inclined towards popularity bias amplification do not necessarily perform worse in terms of utility (NDCG), (3) the majority of the investigated recommenders intensify the popularity bias of the female users.

Markus Reiter-Haas, Emilia Parada-Cabaleiro, Markus Schedl et al.

Providing suitable recommendations is of vital importance to improve the user satisfaction of music recommender systems. Here, users often listen to the same track repeatedly and appreciate recommendations of the same song multiple times. Thus, accounting for users' relistening behavior is critical for music recommender systems. In this paper, we describe a psychology-informed approach to model and predict music relistening behavior that is inspired by studies in music psychology, which relate music preferences to human memory. We adopt a well-established psychological theory of human cognition that models the operations of human memory, i.e., Adaptive Control of Thought-Rational (ACT-R). In contrast to prior work, which uses only the base-level component of ACT-R, we utilize five components of ACT-R, i.e., base-level, spreading, partial matching, valuation, and noise, to investigate the effect of five factors on music relistening behavior: (i) recency and frequency of prior exposure to tracks, (ii) co-occurrence of tracks, (iii) the similarity between tracks, (iv) familiarity with tracks, and (v) randomness in behavior. On a dataset of 1.7 million listening events from Last.fm, we evaluate the performance of our approach by sequentially predicting the next track(s) in user sessions. We find that recency and frequency of prior exposure to tracks is an effective predictor of relistening behavior. Besides, considering the co-occurrence of tracks and familiarity with tracks further improves performance in terms of R-precision. We hope that our work inspires future research on the merits of considering cognitive aspects of memory retrieval to model and predict complex user behavior.

Yashar Deldjoo, Markus Schedl, Peter Knees

The music domain is among the most important ones for adopting recommender systems technology. In contrast to most other recommendation domains, which predominantly rely on collaborative filtering (CF) techniques, music recommenders have traditionally embraced content-based (CB) approaches. In the past years, music recommendation models that leverage collaborative and content data -- which we refer to as content-driven models -- have been replacing pure CF or CB models. In this survey, we review 55 articles on content-driven music recommendation. Based on a thorough literature analysis, we first propose an onion model comprising five layers, each of which corresponds to a category of music content we identified: signal, embedded metadata, expert-generated content, user-generated content, and derivative content. We provide a detailed characterization of each category along several dimensions. Second, we identify six overarching challenges, according to which we organize our main discussion: increasing recommendation diversity and novelty, providing transparency and explanations, accomplishing context-awareness, recommending sequences of music, improving scalability and efficiency, and alleviating cold start. Each article addresses one or more of these challenges is categorized according to the content layers of our onion model, the article's goal(s), and main methodological choices. Furthermore, articles are discussed in temporal order to shed light on the evolution of content-driven music recommendation strategies. Finally, we provide our personal selection of the persisting grand challenges which are still waiting to be solved in future research endeavors.

Oleg Lesota, Navid Rekabsaz, Daniel Cohen et al.

Existing neural ranking models follow the text matching paradigm, where document-to-query relevance is estimated through predicting the matching score. Drawing from the rich literature of classical generative retrieval models, we introduce and formalize the paradigm of deep generative retrieval models defined via the cumulative probabilities of generating query terms. This paradigm offers a grounded probabilistic view on relevance estimation while still enabling the use of modern neural architectures. In contrast to the matching paradigm, the probabilistic nature of generative rankers readily offers a fine-grained measure of uncertainty. We adopt several current neural generative models in our framework and introduce a novel generative ranker (T-PGN), which combines the encoding capacity of Transformers with the Pointer Generator Network model. We conduct an extensive set of evaluation experiments on passage retrieval, leveraging the MS MARCO Passage Re-ranking and TREC Deep Learning 2019 Passage Re-ranking collections. Our results show the significantly higher performance of the T-PGN model when compared with other generative models. Lastly, we demonstrate that exploiting the uncertainty information of deep generative rankers opens new perspectives to query/collection understanding, and significantly improves the cut-off prediction task.

Rosie Jones, Hamed Zamani, Markus Schedl et al.

Podcasts are spoken documents across a wide-range of genres and styles, with growing listenership across the world, and a rapidly lowering barrier to entry for both listeners and creators. The great strides in search and recommendation in research and industry have yet to see impact in the podcast space, where recommendations are still largely driven by word of mouth. In this perspective paper, we highlight the many differences between podcasts and other media, and discuss our perspective on challenges and future research directions in the domain of podcast information access.

Navid Rekabsaz, Simone Kopeinik, Markus Schedl

Societal biases resonate in the retrieved contents of information retrieval (IR) systems, resulting in reinforcing existing stereotypes. Approaching this issue requires established measures of fairness in respect to the representation of various social groups in retrieval results, as well as methods to mitigate such biases, particularly in the light of the advances in deep ranking models. In this work, we first provide a novel framework to measure the fairness in the retrieved text contents of ranking models. Introducing a ranker-agnostic measurement, the framework also enables the disentanglement of the effect on fairness of collection from that of rankers. To mitigate these biases, we propose AdvBert, a ranking model achieved by adapting adversarial bias mitigation for IR, which jointly learns to predict relevance and remove protected attributes. We conduct experiments on two passage retrieval collections (MSMARCO Passage Re-ranking and TREC Deep Learning 2019 Passage Re-ranking), which we extend by fairness annotations of a selected subset of queries regarding gender attributes. Our results on the MSMARCO benchmark show that, (1) all ranking models are less fair in comparison with ranker-agnostic baselines, and (2) the fairness of Bert rankers significantly improves when using the proposed AdvBert models. Lastly, we investigate the trade-off between fairness and utility, showing that we can maintain the significant improvements in fairness without any significant loss in utility.

Navid Rekabsaz, Oleg Lesota, Markus Schedl et al.

Click logs are valuable resources for a variety of information retrieval (IR) tasks. This includes query understanding/analysis, as well as learning effective IR models particularly when the models require large amounts of training data. We release a large-scale domain-specific dataset of click logs, obtained from user interactions of the Trip Database health web search engine. Our click log dataset comprises approximately 5.2 million user interactions collected between 2013 and 2020. We use this dataset to create a standard IR evaluation benchmark -- TripClick -- with around 700,000 unique free-text queries and 1.3 million pairs of query-document relevance signals, whose relevance is estimated by two click-through models. As such, the collection is one of the few datasets offering the necessary data richness and scale to train neural IR models with a large amount of parameters, and notably the first in the health domain. Using TripClick, we conduct experiments to evaluate a variety of IR models, showing the benefits of exploiting this data to train neural architectures. In particular, the evaluation results show that the best performing neural IR model significantly improves the performance by a large margin relative to classical IR models, especially for more frequent queries.

Dominik Kowald, Peter Muellner, Eva Zangerle et al.

Music recommender systems have become an integral part of music streaming services such as Spotify and Last.fm to assist users navigating the extensive music collections offered by them. However, while music listeners interested in mainstream music are traditionally served well by music recommender systems, users interested in music beyond the mainstream (i.e., non-popular music) rarely receive relevant recommendations. In this paper, we study the characteristics of beyond-mainstream music and music listeners and analyze to what extent these characteristics impact the quality of music recommendations provided. Therefore, we create a novel dataset consisting of Last.fm listening histories of several thousand beyond-mainstream music listeners, which we enrich with additional metadata describing music tracks and music listeners. Our analysis of this dataset shows four subgroups within the group of beyond-mainstream music listeners that differ not only with respect to their preferred music but also with their demographic characteristics. Furthermore, we evaluate the quality of music recommendations that these subgroups are provided with four different recommendation algorithms where we find significant differences between the groups. Specifically, our results show a positive correlation between a subgroup's openness towards music listened to by members of other subgroups and recommendation accuracy. We believe that our findings provide valuable insights for developing improved user models and recommendation approaches to better serve beyond-mainstream music listeners.

Markus Schedl, Christine Bauer, Wolfgang Reisinger et al.

Music preferences are strongly shaped by the cultural and socio-economic background of the listener, which is reflected, to a considerable extent, in country-specific music listening profiles. Previous work has already identified several country-specific differences in the popularity distribution of music artists listened to. In particular, what constitutes the "music mainstream" strongly varies between countries. To complement and extend these results, the article at hand delivers the following major contributions: First, using state-of-the-art unsupervised learning techniques, we identify and thoroughly investigate (1) country profiles of music preferences on the fine-grained level of music tracks (in contrast to earlier work that relied on music preferences on the artist level) and (2) country archetypes that subsume countries sharing similar patterns of listening preferences. Second, we formulate four user models that leverage the user's country information on music preferences. Among others, we propose a user modeling approach to describe a music listener as a vector of similarities over the identified country clusters or archetypes. Third, we propose a context-aware music recommendation system that leverages implicit user feedback, where context is defined via the four user models. More precisely, it is a multi-layer generative model based on a variational autoencoder, in which contextual features can influence recommendations through a gating mechanism. Fourth, we thoroughly evaluate the proposed recommendation system and user models on a real-world corpus of more than one billion listening records of users around the world (out of which we use 369 million in our experiments) and show its merits vis-a-vis state-of-the-art algorithms that do not exploit this type of context information.

Navid Rekabsaz, Markus Schedl

Concerns regarding the footprint of societal biases in information retrieval (IR) systems have been raised in several previous studies. In this work, we examine various recent IR models from the perspective of the degree of gender bias in their retrieval results. To this end, we first provide a bias measurement framework which includes two metrics to quantify the degree of the unbalanced presence of gender-related concepts in a given IR model's ranking list. To examine IR models by means of the framework, we create a dataset of non-gendered queries, selected by human annotators. Applying these queries to the MS MARCO Passage retrieval collection, we then measure the gender bias of a BM25 model and several recent neural ranking models. The results show that while all models are strongly biased toward male, the neural models, and in particular the ones based on contextualized embedding models, significantly intensify gender bias. Our experiments also show an overall increase in the gender bias of neural models when they exploit transfer learning, namely when they use (already biased) pre-trained embeddings.

Dominik Kowald, Elisabeth Lex, Markus Schedl

In this paper, we introduce a psychology-inspired approach to model and predict the music genre preferences of different groups of users by utilizing human memory processes. These processes describe how humans access information units in their memory by considering the factors of (i) past usage frequency, (ii) past usage recency, and (iii) the current context. Using a publicly available dataset of more than a billion music listening records shared on the music streaming platform Last.fm, we find that our approach provides significantly better prediction accuracy results than various baseline algorithms for all evaluated user groups, i.e., (i) low-mainstream music listeners, (ii) medium-mainstream music listeners, and (iii) high-mainstream music listeners. Furthermore, our approach is based on a simple psychological model, which contributes to the transparency and explainability of the calculated predictions.

Markus Schedl, Christine Bauer

In this paper, we analyze a large dataset of user-generated music listening events from Last.fm, focusing on users aged 6 to 18 years. Our contribution is two-fold. First, we study the music genre preferences of this young user group and analyze these preferences for homogeneity within more fine-grained age groups and with respect to gender and countries. Second, we investigate the performance of a collaborative filtering recommender when tailoring music recommendations to different age groups. We find that doing so improves performance for all user groups up to 18 years, but decreases performance for adult users aged 19 years and older.

Christine Bauer, Markus Schedl

Popularity-based approaches are widely adopted in music recommendation systems, both in industry and research. However, as the popularity distribution of music items typically is a long-tail distribution, popularity-based approaches to music recommendation fall short in satisfying listeners that have specialized music. The contribution of this article is three-fold. We provide several quantitative measures describing the proximity of a user's music preference to the music mainstream. We define the measures at two levels: relating a listener's music preferences to the global music preferences of all users, or relating them to music preferences of the user's country. Moreover, we adopt a distribution-based and a rank-based approach as means to decrease bias towards the head of the long-tail distribution. We analyze differences between countries in terms of their level of mainstreaminess, uncover both positive and negative outliers (substantially higher and lower country-specific popularity, respectively, compared to the global mainstream), and investigate differences between countries in terms of listening preferences related to popular music artists. We use the standardized LFM-1b dataset, from which we analyze about 8 million listening events shared by about 53,000 users (from 47 countries) of the music streaming platform Last.fm. We show that there are substantial country-specific differences in listeners' music consumption behavior with respect to the most popular artists listened to. We conduct rating prediction experiments in which we tailor recommendations to a user's level of preference for the music mainstream using the proposed 6 mainstreaminess measures. Results suggest that, in terms of rating prediction accuracy, each of the presented mainstreaminess definitions has its merits.

Dominik Kowald, Markus Schedl, Elisabeth Lex

Research has shown that recommender systems are typically biased towards popular items, which leads to less popular items being underrepresented in recommendations. The recent work of Abdollahpouri et al. in the context of movie recommendations has shown that this popularity bias leads to unfair treatment of both long-tail items as well as users with little interest in popular items. In this paper, we reproduce the analyses of Abdollahpouri et al. in the context of music recommendation. Specifically, we investigate three user groups from the LastFM music platform that are categorized based on how much their listening preferences deviate from the most popular music among all LastFM users in the dataset: (i) low-mainstream users, (ii) medium-mainstream users, and (iii) high-mainstream users. In line with Abdollahpouri et al., we find that state-of-the-art recommendation algorithms favor popular items also in the music domain. However, their proposed Group Average Popularity metric yields different results for LastFM than for the movie domain, presumably due to the larger number of available items (i.e., music artists) in the LastFM dataset we use. Finally, we compare the accuracy results of the recommendation algorithms for the three user groups and find that the low-mainstreaminess group significantly receives the worst recommendations.

Luca Luciano Costanzo, Yashar Deldjoo, Maurizio Ferrari Dacrema et al.

In order to improve the accuracy of recommendations, many recommender systems nowadays use side information beyond the user rating matrix, such as item content. These systems build user profiles as estimates of users' interest on content (e.g., movie genre, director or cast) and then evaluate the performance of the recommender system as a whole e.g., by their ability to recommend relevant and novel items to the target user. The user profile modelling stage, which is a key stage in content-driven RS is barely properly evaluated due to the lack of publicly available datasets that contain user preferences on content features of items. To raise awareness of this fact, we investigate differences between explicit user preferences and implicit user profiles. We create a dataset of explicit preferences towards content features of movies, which we release publicly. We then compare the collected explicit user feature preferences and implicit user profiles built via state-of-the-art user profiling models. Our results show a maximum average pairwise cosine similarity of 58.07\% between the explicit feature preferences and the implicit user profiles modelled by the best investigated profiling method and considering movies' genres only. For actors and directors, this maximum similarity is only 9.13\% and 17.24\%, respectively. This low similarity between explicit and implicit preference models encourages a more in-depth study to investigate and improve this important user profile modelling step, which will eventually translate into better recommendations.