Large Language Models (LLMs) have started to demonstrate the ability to persuade humans, yet our understanding of how this dynamic transpires is limited. Recent work has used linear probes, lightweight tools for analyzing model representations, to study various LLM skills such as the ability to model user sentiment and political perspective. Motivated by this, we apply probes to study persuasion dynamics in natural, multi-turn conversations. We leverage insights from cognitive science to train probes on distinct aspects of persuasion: persuasion success, persuadee personality, and persuasion strategy. Despite their simplicity, we show that they capture various aspects of persuasion at both the sample and dataset levels. For instance, probes can identify the point in a conversation where the persuadee was persuaded or where persuasive success generally occurs across the entire dataset. We also show that in addition to being faster than expensive prompting-based approaches, probes can do just as well and even outperform prompting in some settings, such as when uncovering persuasion strategy. This suggests probes as a plausible avenue for studying other complex behaviours such as deception and manipulation, especially in multi-turn settings and large-scale dataset analysis where prompting-based methods would be computationally inefficient.

In the realm of music recommendation, sequential recommenders have shown promise in capturing the dynamic nature of music consumption. A key characteristic of this domain is repetitive listening, where users frequently replay familiar tracks. To capture these repetition patterns, recent research has introduced Personalised Popularity Scores (PPS), which quantify user-specific preferences based on historical frequency. While PPS enhances relevance in recommendation, it often reinforces already-known content, limiting the system's ability to surface novel or serendipitous items - key elements for fostering long-term user engagement and satisfaction. To address this limitation, we build upon RecJPQ, a Transformer-based framework initially developed to improve scalability in large-item catalogues through sub-item decomposition. We repurpose RecJPQ's sub-item architecture to model personalised popularity at a finer granularity. This allows us to capture shared repetition patterns across sub-embeddings - latent structures not accessible through item-level popularity alone. We propose a novel integration of sub-ID-level personalised popularity within the RecJPQ framework, enabling explicit control over the trade-off between accuracy and personalised novelty. Our sub-ID-level PPS method (sPPS) consistently outperforms item-level PPS by achieving significantly higher personalised novelty without compromising recommendation accuracy. Code and experiments are publicly available at https://github.com/sisinflab/Sub-id-Popularity.

The internet is rife with unattributed, deliberately misleading, or otherwise untrustworthy content. Though large language models (LLMs) are often tasked with autonomous web browsing, the extent to which they have learned the simple heuristics human researchers use to navigate this noisy environment is not currently known. In this paper, we introduce the Synthetic Media Literacy Test (SMeL Test), a minimal benchmark that tests the ability of language models to actively filter out untrustworthy information in context. We benchmark a variety of commonly used instruction-tuned LLMs, including reasoning models, and find that no model consistently succeeds; while reasoning in particular is associated with higher scores, even the best API model we test hallucinates up to 70% of the time. Remarkably, larger and more capable models do not necessarily outperform their smaller counterparts. We hope our work sheds more light on this important form of hallucination and guides the development of new methods to combat it.

Large Language Models (LLMs) succeed in many natural language processing tasks. However, their tendency to hallucinate - generate plausible but inconsistent or factually incorrect text - can cause significant problems in certain tasks, including response generation in dialogue. To mitigate this issue, we propose two novel graph knowledge-augmented frameworks, Dialogue Response Generation via Textualised Graphs (TG-DRG) and Graph-Aware Dialogue Response Generation (GA-DRG), which combine reasoning-guided dialogue reformulation, dialogue sense knowledge selection, and graph-enhanced response generation to improve the factuality of dialogue responses. To evaluate the factuality of generated responses, we propose a dialogue fact score that addresses the limitations of existing fact-score methods in dialogue settings, providing a more reliable assessment of factual consistency. We evaluate our methods using different baselines on the OpendialKG and HybriDialogue datasets. Our methods noticeably improve factuality compared to other graph knowledge-augmentation baselines, including the state-of-the-art G-retriever, achieving improvements of 3.47% on OpendialKG and 3.12% on HybriDialogue in terms of dialogue fact score. The code will be released on GitHub.

Creating a meaningful representation by fusing single modalities (e.g., text, images, or audio) is the core concept of multimodal learning. Although several techniques for building multimodal representations have been proven successful, they have not been compared yet. Therefore it has been ambiguous which technique can be expected to yield the best results in a given scenario and what factors should be considered while choosing such a technique. This paper explores the most common techniques for building multimodal data representations -- the late fusion, the early fusion, and the sketch, and compares them in classification tasks. Experiments are conducted on three datasets: Amazon Reviews, MovieLens25M, and MovieLens1M datasets. In general, our results confirm that multimodal representations are able to boost the performance of unimodal models from 0.919 to 0.969 of accuracy on Amazon Reviews and 0.907 to 0.918 of AUC on MovieLens25M. However, experiments on both MovieLens datasets indicate the importance of the meaningful input data to the given task. In this article, we show that the choice of the technique for building multimodal representation is crucial to obtain the highest possible model's performance, that comes with the proper modalities combination. Such choice relies on: the influence that each modality has on the analyzed machine learning (ML) problem; the type of the ML task; the memory constraints while training and predicting phase.

Hate speech is a widespread and harmful form of online discourse, encompassing slurs and defamatory posts that can have serious social, psychological, and sometimes physical impacts on targeted individuals and communities. As social media platforms such as X (formerly Twitter), Facebook, Instagram, Reddit, and others continue to facilitate widespread communication, they also become breeding grounds for hate speech, which has increasingly been linked to real-world hate crimes. Addressing this issue requires the development of robust automated methods to detect hate speech in diverse social media environments. Deep learning approaches, such as vanilla recurrent neural networks (RNNs), long short-term memory (LSTM), and convolutional neural networks (CNNs), have achieved good results, but are often limited by issues such as long-term dependencies and inefficient parallelization. This study represents the comprehensive exploration of transformer-based models for hate speech detection using the MetaHate dataset--a meta-collection of 36 datasets with 1.2 million social media samples. We evaluate multiple state-of-the-art transformer models, including BERT, RoBERTa, GPT-2, and ELECTRA, with fine-tuned ELECTRA achieving the highest performance (F1 score: 0.8980). We also analyze classification errors, revealing challenges with sarcasm, coded language, and label noise.

--Open-vocabulary keyword spotting (KWS) refers to the task of detecting words or terms within speech recordings, regardless of whether they were included in the training data. This paper introduces an open-vocabulary keyword spotting model with state-of-the-art detection accuracy for small-footprint devices. The model is composed of a speech encoder, a target keyword encoder, and a detection network. The speech encoder is either a tiny Whisper or a tiny Conformer . The target keyword encoder is implemented as a hyper-network that takes the desired keyword as a character string and generates a unique set of weights for a convolutional layer, which can be considered as a keyword-specific matched filter . The detection network uses the matched-filter weights to perform a keyword-specific convolution, which guides the cross-attention mechanism of a Perceiver module in determining whether the target term appears in the recording. The results indicate that our system achieves state-of-the-art detection performance and generalizes effectively to out-of-domain conditions, including second-language (L2) speech. Notably, our smallest model, with just 4.2 million parameters, matches or outperforms models that are several times larger, demonstrating both efficiency and robustness. Keyword Spotting (KWS) is the task of identifying a predefined set of target keywords. It is an essential task in speech recognition systems, particularly crucial for voice assistants on small-footprint devices, such as smartphones and small speakers [1]. Throughout the years, various techniques have been explored for this task, beginning with basic deep neural networks based on Convolutional Neural Networks (CNNs) [2], [3] to Recurrent Neural Networks (RNNs) as audio encoders [4], [5]. V arious approaches have been proposed to address this challenge.

Emotions critically influence mental health, driving interest in music-based affective computing via neurophysiological signals with Brain-computer Interface techniques. While prior studies leverage music's accessibility for emotion induction, three key limitations persist: \textbf{(1) Stimulus Constraints}: Music stimuli are confined to small corpora due to copyright and curation costs, with selection biases from heuristic emotion-music mappings that ignore individual affective profiles. \textbf{(2) Modality Specificity}: Overreliance on unimodal neural data (e.g., EEG) ignores complementary insights from cross-modal signal fusion.\textbf{ (3) Portability Limitation}: Cumbersome setups (e.g., 64+ channel gel-based EEG caps) hinder real-world applicability due to procedural complexity and portability barriers. To address these limitations, we propose MEEtBrain, a portable and multimodal framework for emotion analysis (valence/arousal), integrating AI-generated music stimuli with synchronized EEG-fNIRS acquisition via a wireless headband. By MEEtBrain, the music stimuli can be automatically generated by AI on a large scale, eliminating subjective selection biases while ensuring music diversity. We use our developed portable device that is designed in a lightweight headband-style and uses dry electrodes, to simultaneously collect EEG and fNIRS recordings. A 14-hour dataset from 20 participants was collected in the first recruitment to validate the framework's efficacy, with AI-generated music eliciting target emotions (valence/arousal). We are actively expanding our multimodal dataset (44 participants in the latest dataset) and make it publicly available to promote further research and practical applications. \textbf{The dataset is available at https://zju-bmi-lab.github.io/ZBra.

Recently, the information content (IC) of predictions from a Generative Infinite-Vocabulary Transformer (GIVT) has been used to model musical expectancy and surprisal in audio. We investigate the effectiveness of such modelling using IC calculated with autoregressive diffusion models (ADMs). We empirically show that IC estimates of models based on two different diffusion ordinary differential equations (ODEs) describe diverse data better, in terms of negative log-likelihood, than a GIVT. We evaluate diffusion model IC's effectiveness in capturing surprisal aspects by examining two tasks: (1) capturing monophonic pitch surprisal, and (2) detecting segment boundaries in multi-track audio. In both tasks, the diffusion models match or exceed the performance of a GIVT. We hypothesize that the surprisal estimated at different diffusion process noise levels corresponds to the surprisal of music and audio features present at different audio granularities. Testing our hypothesis, we find that, for appropriate noise levels, the studied musical surprisal tasks' results improve. Code is provided on github.com/SonyCSLParis/audioic.

Humans often rely on underlying structural patterns-schemas-to create, whether by writing stories, designing software, or composing music. Schemas help organize ideas and guide exploration, but they are often difficult to discover and apply, especially in complex or unfamiliar domains. My Ph.D. research develops a framework for human-AI schema discovery and application to support creative problem solving. I design systems that support users in sensemaking over examples to abstract schemas, and in operationalizing schemas into human-AI co-creative workflows for application. This research offers insights into how schema-guided interaction can make implicit knowledge more accessible and actionable, advancing more transparent and collaborative human-AI systems.