Content moderation plays a critical role in shaping safe and inclusive online environments, balancing platform standards, user expectations, and regulatory frameworks. Traditionally, this process involves operationalising policies into guidelines, which are then used by downstream human moderators for enforcement, or to further annotate datasets for training machine learning moderation models. However, recent advancements in large language models (LLMs) are transforming this landscape. These models can now interpret policies directly as textual inputs, eliminating the need for extensive data curation. This approach offers unprecedented flexibility, as moderation can be dynamically adjusted through natural language interactions. This paradigm shift raises important questions about how policies are operationalised and the implications for content moderation practices. In this paper, we formalise the emerging policy-as-prompt framework and identify five key challenges across four domains: Technical Implementation (1. translating policy to prompts, 2. sensitivity to prompt structure and formatting), Sociotechnical (3. the risk of technological determinism in policy formation), Organisational (4. evolving roles between policy and machine learning teams), and Governance (5. model governance and accountability). Through analysing these challenges across technical, sociotechnical, organisational, and governance dimensions, we discuss potential mitigation approaches. This research provides actionable insights for practitioners and lays the groundwork for future exploration of scalable and adaptive content moderation systems in digital ecosystems.

Slate generation is a common task in streaming and e-commerce platforms, where multiple items are presented together as a list or ``slate''. Traditional systems focus mostly on item-level ranking and often fail to capture the coherence of the slate as a whole. A key challenge lies in the combinatorial nature of selecting multiple items jointly. To manage this, conventional approaches often assume users interact with only one item at a time, assumption that breaks down when items are meant to be consumed together. In this paper, we introduce DMSG, a generative framework based on diffusion models for prompt-conditioned slate generation. DMSG learns high-dimensional structural patterns and generates coherent, diverse slates directly from natural language prompts. Unlike retrieval-based or autoregressive models, DMSG models the joint distribution over slates, enabling greater flexibility and diversity. We evaluate DMSG in two key domains: music playlist generation and e-commerce bundle creation. In both cases, DMSG produces high-quality slates from textual prompts without explicit personalization signals. Offline and online results show that DMSG outperforms strong baselines in both relevance and diversity, offering a scalable, low-latency solution for prompt-driven recommendation. A live A/B test on a production playlist system further demonstrates increased user engagement and content diversity.

With the breakthroughs in large language models (LLMs), query generation techniques that expand documents and queries with related terms are becoming increasingly popular in the information retrieval field. Such techniques have been shown to improve the effectiveness of traditional lexical retrieval methods by dealing with the vocabulary mismatch problem. Recent work has found that generating queries with a greedy decoding strategy can produce sub-optimal queries, including hallucinations, and proposed to filter out queries before expansion. This `generate-then-filter' approach is costly, as it requires generating multiple queries and applying a relevance model to all of them and does not teach the LLM which of the generated queries is more effective for expansion. To overcome such limitations, we propose Aligned Query Expansion (AQE), a novel approach to enhance query expansion for passage retrieval in open-domain question answering. AQE leverages recent techniques in LLM alignment to fine-tune models for generating query expansions that directly optimize the effectiveness of the retrieval task, eliminating the need for additional filtering steps. This alignment ensures that queries are more relevant, reducing computational costs while improving retrieval effectiveness. Empirical evaluations show that AQE outperforms baseline models for query expansion in both in-domain and out-of-domain settings, demonstrating significant improvements in retrieval effectiveness.

Evaluating personalized recommendations remains a central challenge, especially in long-form audio domains like podcasts, where traditional offline metrics suffer from exposure bias and online methods such as A/B testing are costly and operationally constrained. In this paper, we propose a novel framework that leverages Large Language Models (LLMs) as offline judges to assess the quality of podcast recommendations in a scalable and interpretable manner. Our two-stage profile-aware approach first constructs natural-language user profiles distilled from 90 days of listening history. These profiles summarize both topical interests and behavioral patterns, serving as compact, interpretable representations of user preferences. Rather than prompting the LLM with raw data, we use these profiles to provide high-level, semantically rich context-enabling the LLM to reason more effectively about alignment between a user's interests and recommended episodes. This reduces input complexity and improves interpretability. The LLM is then prompted to deliver fine-grained pointwise and pairwise judgments based on the profile-episode match. In a controlled study with 47 participants, our profile-aware judge matched human judgments with high fidelity and outperformed or matched a variant using raw listening histories. The framework enables efficient, profile-aware evaluation for iterative testing and model selection in recommender systems.

Automatic Music Transcription (AMT) has been recognized as a key enabling technology with a wide range of applications. Given the task's complexity, best results have typically been reported for systems focusing on specific settings, e.g. instrument-specific systems tend to yield improved results over instrument-agnostic methods. Similarly, higher accuracy can be obtained when only estimating frame-wise $f_0$ values and neglecting the harder note event detection. Despite their high accuracy, such specialized systems often cannot be deployed in the real-world. Storage and network constraints prohibit the use of multiple specialized models, while memory and run-time constraints limit their complexity. In this paper, we propose a lightweight neural network for musical instrument transcription, which supports polyphonic outputs and generalizes to a wide variety of instruments (including vocals). Our model is trained to jointly predict frame-wise onsets, multipitch and note activations, and we experimentally show that this multi-output structure improves the resulting frame-level note accuracy. Despite its simplicity, benchmark results show our system's note estimation to be substantially better than a comparable baseline, and its frame-level accuracy to be only marginally below those of specialized state-of-the-art AMT systems. With this work we hope to encourage the community to further investigate low-resource, instrument-agnostic AMT systems.

Music has a unique and complex structure which is challenging for both expert humans and existing AI systems to understand, and presents unique challenges relative to other forms of audio. We present LLark, an instruction-tuned multimodal model for \emph{music} understanding. We detail our process for dataset creation, which involves augmenting the annotations of diverse open-source music datasets and converting them to a unified instruction-tuning format. We propose a multimodal architecture for LLark, integrating a pretrained generative model for music with a pretrained language model. In evaluations on three types of tasks (music understanding, captioning, reasoning), we show that LLark matches or outperforms existing baselines in music understanding, and that humans show a high degree of agreement with its responses in captioning and reasoning tasks. LLark is trained entirely from open-source music data and models, and we make our training code available along with the release of this paper. Additional results and audio examples are at this https URL, and our source code is available at this https URL .

We introduce Feasible Learning (FL), a sample-centric learning paradigm where models are trained by solving a feasibility problem that bounds the loss for each training sample. In contrast to the ubiquitous Empirical Risk Minimization (ERM) framework, which optimizes for average performance, FL demands satisfactory performance on every individual data point. Since any model that meets the prescribed performance threshold is a valid FL solution, the choice of optimization algorithm and its dynamics play a crucial role in shaping the properties of the resulting solutions. In particular, we study a primal-dual approach which dynamically re-weights the importance of each sample during training. To address the challenge of setting a meaningful threshold in practice, we introduce a relaxation of FL that incorporates slack variables of minimal norm. Our empirical analysis, spanning image classification, age regression, and preference optimization in large language models, demonstrates that models trained via FL can learn from data while displaying improved tail behavior compared to ERM, with only a marginal impact on average performance.

Embedding models trained separately on similar data often produce representations that encode stable information but are not directly interchangeable. This lack of interoperability raises challenges in several practical applications, such as model retraining, partial model upgrades, and multimodal search. Driven by these challenges, we study when two sets of embeddings can be aligned by an orthogonal transformation. We show that if pairwise dot products are approximately preserved, then there exists an isometry that closely aligns the two sets, and we provide a tight bound on the alignment error. This insight yields a simple alignment recipe, Procrustes post-processing, that makes two embedding models interoperable while preserving the geometry of each embedding space. Empirically, we demonstrate its effectiveness in three applications: maintaining compatibility across retrainings, combining different models for text retrieval, and improving mixed-modality search, where it achieves state-of-the-art performance.

We present a novel framework for user representation in large-scale recommender systems, aiming at effectively representing diverse user taste in a generalized manner. Our approach employs a two-stage methodology combining representation learning and transfer learning. The representation learning model uses an autoencoder that compresses various user features into a representation space. In the second stage, downstream task-specific models leverage user representations via transfer learning instead of curating user features individually. We further augment this methodology on the representation's input features to increase flexibility and enable reaction to user events, including new user experiences, in Near-Real Time. Additionally, we propose a novel solution to manage deployment of this framework in production models, allowing downstream models to work independently. We validate the performance of our framework through rigorous offline and online experiments within a large-scale system, showcasing its remarkable efficacy across multiple evaluation tasks. Finally, we show how the proposed framework can significantly reduce infrastructure costs compared to alternative approaches.

We present a novel podcast recommender system deployed at industrial scale. This system successfully optimizes personal listening journeys that unfold over months for hundreds of millions of listeners. In deviating from the pervasive industry practice of optimizing machine learning algorithms for short-term proxy metrics, the system substantially improves long-term performance in A/B tests. The paper offers insights into how our methods cope with attribution, coordination, and measurement challenges that usually hinder such long-term optimization. To contextualize these practical insights within a broader academic framework, we turn to reinforcement learning (RL). Using the language of RL, we formulate a comprehensive model of users' recurring relationships with a recommender system. Then, within this model, we identify our approach as a policy improvement update to a component of the existing recommender system, enhanced by tailored modeling of value functions and user-state representations. Illustrative offline experiments suggest this specialized modeling reduces data requirements by as much as a factor of 120,000 compared to black-box approaches.

Sequential recommendation models the dynamics of a user's previous behaviors in order to forecast the next item, and has drawn a lot of attention. Transformer-based approaches, which embed items as vectors and use dot-product self-attention to measure the relationship between items, demonstrate superior capabilities among existing sequential methods. However, users' real-world sequential behaviors are \textit{\textbf{uncertain}} rather than deterministic, posing a significant challenge to present techniques. We further suggest that dot-product-based approaches cannot fully capture \textit{\textbf{collaborative transitivity}}, which can be derived in item-item transitions inside sequences and is beneficial for cold start items. We further argue that BPR loss has no constraint on positive and sampled negative items, which misleads the optimization. We propose a novel \textbf{STO}chastic \textbf{S}elf-\textbf{A}ttention~(STOSA) to overcome these issues. STOSA, in particular, embeds each item as a stochastic Gaussian distribution, the covariance of which encodes the uncertainty. We devise a novel Wasserstein Self-Attention module to characterize item-item position-wise relationships in sequences, which effectively incorporates uncertainty into model training. Wasserstein attentions also enlighten the collaborative transitivity learning as it satisfies triangle inequality. Moreover, we introduce a novel regularization term to the ranking loss, which assures the dissimilarity between positive and the negative items. Extensive experiments on five real-world benchmark datasets demonstrate the superiority of the proposed model over state-of-the-art baselines, especially on cold start items. The code is available in \url{https://github.com/zfan20/STOSA}.

This study presents a systematic comparison of methods for individual treatment assignment, a general problem that arises in many applications and has received significant attention from economists, computer scientists, and social scientists. We group the various methods proposed in the literature into three general classes of algorithms (or metalearners): learning models to predict outcomes (the O-learner), learning models to predict causal effects (the E-learner), and learning models to predict optimal treatment assignments (the A-learner). We compare the metalearners in terms of (1) their level of generality and (2) the objective function they use to learn models from data; we then discuss the implications that these characteristics have for modeling and decision making. Notably, we demonstrate analytically and empirically that optimizing for the prediction of outcomes or causal effects is not the same as optimizing for treatment assignments, suggesting that in general the A-learner should lead to better treatment assignments than the other metalearners. We demonstrate the practical implications of our findings in the context of choosing, for each user, the best algorithm for playlist generation in order to optimize engagement. This is the first comparison of the three different metalearners on a real-world application at scale (based on more than half a billion individual treatment assignments). In addition to supporting our analytical findings, the results show how large A/B tests can provide substantial value for learning treatment assignment policies, rather than simply choosing the variant that performs best on average.

Lyrics alignment gained considerable attention in recent years. State-of-the-art systems either re-use established speech recognition toolkits, or design end-to-end solutions involving a Connectionist Temporal Classification (CTC) loss. However, both approaches suffer from specific weaknesses: toolkits are known for their complexity, and CTC systems use a loss designed for transcription which can limit alignment accuracy. In this paper, we use instead a contrastive learning procedure that derives cross-modal embeddings linking the audio and text domains. This way, we obtain a novel system that is simple to train end-to-end, can make use of weakly annotated training data, jointly learns a powerful text model, and is tailored to alignment. The system is not only the first to yield an average absolute error below 0.2 seconds on the standard Jamendo dataset but it is also robust to other languages, even when trained on English data only. Finally, we release word-level alignments for the JamendoLyrics Multi-Lang dataset.

Chain-of-thought (CoT) traces promise transparency for reasoning language models, but prior work shows they are not always faithful reflections of internal computation. This raises challenges for oversight: practitioners may misinterpret decorative reasoning as genuine. We introduce Concept Walk, a general framework for tracing how a model's internal stance evolves with respect to a concept direction during reasoning. Unlike surface text, Concept Walk operates in activation space, projecting each reasoning step onto the concept direction learned from contrastive data. This allows us to observe whether reasoning traces shape outcomes or are discarded. As a case study, we apply Concept Walk to the domain of Safety using Qwen 3-4B. We find that in 'easy' cases, perturbed CoTs are quickly ignored, indicating decorative reasoning, whereas in 'hard' cases, perturbations induce sustained shifts in internal activations, consistent with faithful reasoning. The contribution is methodological: Concept Walk provides a lens to re-examine faithfulness through concept-specific internal dynamics, helping identify when reasoning traces can be trusted and when they risk misleading practitioners.

Detecting whether an LLM hallucinates is an important research challenge. One promising way of doing so is to estimate the semantic entropy (Farquhar et al., 2024) of the distribution of generated sequences. We propose a new algorithm for doing that, with two main advantages. First, due to us taking the Bayesian approach, we achieve a much better quality of semantic entropy estimates for a given budget of samples from the LLM. Second, we are able to tune the number of samples adaptively so that `harder' contexts receive more samples. We demonstrate empirically that our approach systematically beats the baselines, requiring only 53% of samples used by Farquhar et al. (2024) to achieve the same quality of hallucination detection as measured by AUROC. Moreover, quite counterintuitively, our estimator is useful even with just one sample from the LLM.