England's players were racially abused during their second Test victory over Argentina in San Juan on 12 July. Team officials lodged a complaint to governing body World Rugby over the incident that occurred when the visitors' replacements were warming up in the first half. "While it is clear that an incident took place, we regret that the individuals responsible could not be identified," said World Rugby, adding their investigation included witness statements and video analysis. "Intense efforts were made to identify the small group of five or seven individuals responsible within a crowd of over 20,000 spectators," said Gabriel Travaglini, president of the Union Argentina de Rugby (UAR). "Unfortunately, despite an exhaustive search, it was not possible to identify the perpetrators. "We strongly condemn all acts of racism and stand in solidarity with the England rugby players who felt aggrieved." He added that the UAR would work with World Rugby to educate fans. There have been several recent high-profile cases of discriminatory behaviour in Argentine sport. In 2020, Pablo Matera and Guido Petti, both of whom played in the match in San Juan, were suspended from the team after racist remarks they had made on social media several years earlier were unearthed. In 2024, Chelsea footballer Enzo Fernandez apologised to team-mates after being filmed joining in with a chant that questioned the heritage of France's black and mixed race players. "Rugby completely condemns discriminatory behaviour of any kind," said World Rugby chairman Brett Robinson. "We offer our full support to the players involved and want them to know that rugby stands with them in opposing racism.

We develop a semi-amortized, policy-based, approach to Bayesian experimental design (BED) called Stepwise Deep Adaptive Design (Step-DAD). Like existing, fully amortized, policy-based BED approaches, Step-DAD trains a design policy upfront before the experiment. However, rather than keeping this policy fixed, Step-DAD periodically updates it as data is gathered, refining it to the particular experimental instance. This test-time adaptation improves both the flexibility and the robustness of the design strategy compared with existing approaches. Empirically, Step-DAD consistently demonstrates superior decision-making and robustness compared with current state-of-the-art BED methods.

Causal forests have become a popular tool for estimating how treatment effects vary across individuals (Wager and Athey, 2018). They are used in a growing number of domains--including marketing, operations, economics, and public policy--to personalize interventions and inform targeting strategies. Since 2019, dozens of papers in INFORMS journals alone have applied causal forests to experimental or observational data (see Appendix C), often with the goal of estimating individual-level treatment effects. The method builds on a familiar idea: instead of estimating a single average effect for the whole population, we split the population into subgroups based on observed features and estimate effects within each group. This is conceptually similar to how random forests estimate outcomes, except now the goal is to estimate causal effects. But there is a crucial modeling difference: unlike random forests, which typically use the full training data for both splitting and estimation, causal forests often divide the training data in two--using one part to decide how to form the subgroups, and the other to estimate effects within them. This practice, known as honest estimation, is meant to prevent overfitting and selection bias (Athey and Imbens, 2016). It is the default in widely used software packages such as grf (Athey et al., 2019) and EconML (Battocchi et al., 2019), and is commonly recommended in applied research. But is this default always a good idea? 1

Recent work has shown that dynamic programming (DP) methods for finding static CVaR-optimal policies in Markov Decision Processes (MDPs) can fail when based on the dual formulation, yet the root cause for the failure has remained unclear. We expand on these findings by shifting focus from policy optimization to the seemingly simpler task of policy evaluation. We show that evaluating the static CVaR of a given policy can be framed as two distinct minimization problems. For their solutions to match, a set of ``risk-assignment consistency constraints'' must be satisfied, and we demonstrate that the intersection of the constraints being empty is the source of previously observed evaluation errors. Quantifying the evaluation error as the CVaR evaluation gap, we then demonstrate that the issues observed when optimizing over the dual-based CVaR DP are explained by the returned policy having a non-zero CVaR evaluation gap. We then leverage our proposed risk-assignment perspective to prove that the search for a single, uniformly optimal policy via on the dual CVaR decomposition is fundamentally limited, identifying an MDP where no single policy can be optimal across all initial risk levels.

Large Language Models (LLMs) frequently reproduce the gender- and sexual-identity prejudices embedded in their training corpora, leading to outputs that marginalize LGBTQIA+ users. Hence, reducing such biases is of great importance. To achieve this, we evaluate two parameter-efficient fine-tuning (PEFT) techniques - Low-Rank Adaptation (LoRA) and soft-prompt tuning - as lightweight alternatives to full-model fine-tuning for mitigating such biases. Using the WinoQueer benchmark, we quantify bias in three open-source LLMs and observe baseline bias scores reaching up to 98 (out of 100) across a range of queer identities defined by gender and/or sexual orientation, where 50 would indicate neutrality. Fine-tuning with LoRA (< 0.1% additional parameters) on a curated QueerNews corpus reduces those scores by up to 50 points and raises neutrality from virtually 0% to as much as 36%. Soft-prompt tuning (10 virtual tokens) delivers only marginal improvements. These findings show that LoRA can deliver meaningful fairness gains with minimal computation. We advocate broader adoption of community-informed PEFT, the creation of larger queer-authored corpora, and richer evaluation suites beyond WinoQueer, coupled with ongoing audits to keep LLMs inclusive.

This study examines the role of human judges in legal decision-making by using machine learning to predict child physical custody outcomes in French appellate courts. Building on the legal realism-formalism debate, we test whether individual judges' decision-making patterns significantly influence case outcomes, challenging the assumption that judges are neutral variables that apply the law uniformly. To ensure compliance with French privacy laws, we implement a strict pseudonymization process. Our analysis uses 18,937 living arrangements rulings extracted from 10,306 cases. We compare models trained on individual judges' past rulings (specialist models) with a judge-agnostic model trained on aggregated data (generalist models). The prediction pipeline is a hybrid approach combining large language models (LLMs) for structured feature extraction and ML models for outcome prediction (RF, XGB and SVC). Our results show that specialist models consistently achieve higher predictive accuracy than the general model, with top-performing models reaching F1 scores as high as 92.85%, compared to the generalist model's 82.63% trained on 20x to 100x more samples. Specialist models capture stable individual patterns that are not transferable to other judges. In-Domain and Cross-Domain validity tests provide empirical support for legal realism, demonstrating that judicial identity plays a measurable role in legal outcomes. All data and code used will be made available.

Point of interest (POI) recommendation can play a pivotal role in enriching tourists' experiences by suggesting context-dependent and preference-matching locations and activities, such as restaurants, landmarks, itineraries, and cultural attractions. Unlike some more common recommendation domains (e.g., music and video), POI recommendation is inherently high-stakes: users invest significant time, money, and effort to search, choose, and consume these suggested POIs. Despite the numerous research works in the area, several fundamental issues remain unresolved, hindering the real-world applicability of the proposed approaches. In this paper, we discuss the current status of the POI recommendation problem and the main challenges we have identified. The first contribution of this paper is a critical assessment of the current state of POI recommendation research and the identification of key shortcomings across three main dimensions: datasets, algorithms, and evaluation methodologies. We highlight persistent issues such as the lack of standardized benchmark datasets, flawed assumptions in the problem definition and model design, and inadequate treatment of biases in the user behavior and system performance. The second contribution is a structured research agenda that, starting from the identified issues, introduces important directions for future work related to multistakeholder design, context awareness, data collection, trustworthiness, novel interactions, and real-world evaluation.

The integration of agential artificial intelligence into socioeconomic systems requires us to reexamine the evolutionary processes that describe changes in our economic institutions. This article synthesizes three frameworks: multi-level selection theory, Aoki's view of firms as computational processes, and Ostrom's design principles for robust institutions. We develop a framework where selection operates concurrently across organizational levels, firms implement distributed inference via game-theoretic architectures, and Ostrom-style rules evolve as alignment mechanisms that address AI-related risks. This synthesis yields a multi-level Price equation expressed over nested games, providing quantitative metrics for how selection and governance co-determine economic outcomes. We examine connections to Acemoglu's work on inclusive institutions, analyze how institutional structures shape AI deployment, and demonstrate the framework's explanatory power via case studies. We conclude by proposing a set of design principles that operationalize alignment between humans and AI across institutional layers, enabling scalable, adaptive, and inclusive governance of agential AI systems. We conclude with practical policy recommendations and further research to extend these principles into real-world implementation.

This work is still in progress Russian speech synthesis presents distinctive challenges, including vowel reduction, consonant devoicing, variable stress patterns, homograph ambiguity, and unnatural intonation. This paper introduces Balalaika, a novel dataset comprising more than 2,000 hours of studio-quality Russian speech with comprehensive textual annotations, including punctuation and stress markings. Experimental results show that models trained on Balalaika significantly outperform those trained on existing datasets in both speech synthesis and enhancement tasks.

Recent advances in generative models have demonstrated remarkable capabilities in producing high-quality images, but their reliance on large-scale unlabeled data has raised significant safety and copyright concerns. Efforts to address these issues by erasing unwanted concepts have shown promise. However, many existing erasure methods involve excessive modifications that compromise the overall utility of the model. In this work, we address these issues by formulating a novel minimalist concept erasure objective based \emph{only} on the distributional distance of final generation outputs. Building on our formulation, we derive a tractable loss for differentiable optimization that leverages backpropagation through all generation steps in an end-to-end manner. We also conduct extensive analysis to show theoretical connections with other models and methods. To improve the robustness of the erasure, we incorporate neuron masking as an alternative to model fine-tuning. Empirical evaluations on state-of-the-art flow-matching models demonstrate that our method robustly erases concepts without degrading overall model performance, paving the way for safer and more responsible generative models.