Golnoosh Farnadi

Aditi Khandelwal, Marius Mosbach, Verna Dankers et al.

Mixture-of-Experts (MoE) models are widely used to scale language models, yet their expert routing behavior and adaptation in a multilingual setting remain underexplored. In this work, we study multilingual routing dynamics during continual pre-training of an English-centric MoE model on a multilingual corpus, analyzing how expert usage varies across languages. We find that continual multilingual pre-training leads to diffused, language-agnostic routing in early and middle layers, with language specialization primarily emerging in the final layers. We also show that token-level vocabulary overlap between languages plays an important role in how languages are routed. Motivated by these findings, we propose a parameter-efficient adaptation strategy that updates language-specific and shared experts in the final MoE layers. Experiments on MultiBLiMP and Belebele show that our method achieves a strong performance-efficiency trade-off, attaining competitive performance relative to fine-tuning complete final layers, while updating less than 2% of the parameters. Overall, our findings provide insights into where and how language specialization emerges in MoEs during continual pre-training and provide practical insights for low-resource multilingual adaptation. Our code is available at https://github.com/aditi184/moe-routing-adaptation.

Meraj Hashemizadeh, Juan Ramirez, Rohan Sukumaran et al. · mila

Model pruning is a popular approach to enable the deployment of large deep learning models on edge devices with restricted computational or storage capacities. Although sparse models achieve performance comparable to that of their dense counterparts at the level of the entire dataset, they exhibit high accuracy drops for some data sub-groups. Existing methods to mitigate this disparate impact induced by pruning (i) rely on surrogate metrics that address the problem indirectly and have limited interpretability; or (ii) scale poorly with the number of protected sub-groups in terms of computational cost. We propose a constrained optimization approach that directly addresses the disparate impact of pruning: our formulation bounds the accuracy change between the dense and sparse models, for each sub-group. This choice of constraints provides an interpretable success criterion to determine if a pruned model achieves acceptable disparity levels. Experimental results demonstrate that our technique scales reliably to problems involving large models and hundreds of protected sub-groups.

Sikha Pentyala, Nicola Neophytou, Anderson Nascimento et al. · uw

Group fairness ensures that the outcome of machine learning (ML) based decision making systems are not biased towards a certain group of people defined by a sensitive attribute such as gender or ethnicity. Achieving group fairness in Federated Learning (FL) is challenging because mitigating bias inherently requires using the sensitive attribute values of all clients, while FL is aimed precisely at protecting privacy by not giving access to the clients' data. As we show in this paper, this conflict between fairness and privacy in FL can be resolved by combining FL with Secure Multiparty Computation (MPC) and Differential Privacy (DP). In doing so, we propose a method for training group-fair ML models in cross-device FL under complete and formal privacy guarantees, without requiring the clients to disclose their sensitive attribute values.

Ahmad-Reza Ehyaei, Kiarash Mohammadi, Amir-Hossein Karimi et al. · eth-zurich

As responsible AI gains importance in machine learning algorithms, properties such as fairness, adversarial robustness, and causality have received considerable attention in recent years. However, despite their individual significance, there remains a critical gap in simultaneously exploring and integrating these properties. In this paper, we propose a novel approach that examines the relationship between individual fairness, adversarial robustness, and structural causal models in heterogeneous data spaces, particularly when dealing with discrete sensitive attributes. We use causal structural models and sensitive attributes to create a fair metric and apply it to measure semantic similarity among individuals. By introducing a novel causal adversarial perturbation and applying adversarial training, we create a new regularizer that combines individual fairness, causality, and robustness in the classifier. Our method is evaluated on both real-world and synthetic datasets, demonstrating its effectiveness in achieving an accurate classifier that simultaneously exhibits fairness, adversarial robustness, and causal awareness.

Rebecca Salganik, Fernando Diaz, Golnoosh Farnadi

As online music platforms grow, music recommender systems play a vital role in helping users navigate and discover content within their vast musical databases. At odds with this larger goal, is the presence of popularity bias, which causes algorithmic systems to favor mainstream content over, potentially more relevant, but niche items. In this work we explore the intrinsic relationship between music discovery and popularity bias. To mitigate this issue we propose a domain-aware, individual fairness-based approach which addresses popularity bias in graph neural network (GNNs) based recommender systems. Our approach uses individual fairness to reflect a ground truth listening experience, i.e., if two songs sound similar, this similarity should be reflected in their representations. In doing so, we facilitate meaningful music discovery that is robust to popularity bias and grounded in the music domain. We apply our BOOST methodology to two discovery based tasks, performing recommendations at both the playlist level and user level. Then, we ground our evaluation in the cold start setting, showing that our approach outperforms existing fairness benchmarks in both performance and recommendation of lesser-known content. Finally, our analysis explains why our proposed methodology is a novel and promising approach to mitigating popularity bias and improving the discovery of new and niche content in music recommender systems.

Niloofar Mireshghallah, Maria Antoniak, Yash More et al.

Measuring personal disclosures made in human-chatbot interactions can provide a better understanding of users' AI literacy and facilitate privacy research for large language models (LLMs). We run an extensive, fine-grained analysis on the personal disclosures made by real users to commercial GPT models, investigating the leakage of personally identifiable and sensitive information. To understand the contexts in which users disclose to chatbots, we develop a taxonomy of tasks and sensitive topics, based on qualitative and quantitative analysis of naturally occurring conversations. We discuss these potential privacy harms and observe that: (1) personally identifiable information (PII) appears in unexpected contexts such as in translation or code editing (48% and 16% of the time, respectively) and (2) PII detection alone is insufficient to capture the sensitive topics that are common in human-chatbot interactions, such as detailed sexual preferences or specific drug use habits. We believe that these high disclosure rates are of significant importance for researchers and data curators, and we call for the design of appropriate nudging mechanisms to help users moderate their interactions.

Maryam Molamohammadi, Afaf Taik, Nicolas Le Roux et al.

The growing utilization of machine learning (ML) in decision-making processes raises questions about its benefits to society. In this study, we identify and analyze three axes of heterogeneity that significantly influence the trajectory of ML products. These axes are i) values, culture and regulations, ii) data composition, and iii) resource and infrastructure capacity. We demonstrate how these axes are interdependent and mutually influence one another, emphasizing the need to consider and address them jointly. Unfortunately, the current research landscape falls short in this regard, often failing to adopt a holistic approach. We examine the prevalent practices and methodologies that skew these axes in favor of a selected few, resulting in power concentration, homogenized control, and increased dependency. We discuss how this fragmented study of the three axes poses a significant challenge, leading to an impractical solution space that lacks reflection of real-world scenarios. Addressing these issues is crucial to ensure a more comprehensive understanding of the interconnected nature of society and to foster the democratic and inclusive development of ML systems that are more aligned with real-world complexities and its diverse requirements.

Alireza Dehghanpour Farashah, Aditi Khandelwal, Marylou Fauchard et al. · microsoft-research

As multilingual large language models become more widely used, ensuring their safety and fairness across diverse linguistic contexts presents unique challenges. While existing research on machine unlearning has primarily focused on monolingual settings, typically English, multilingual environments introduce additional complexities due to cross-lingual knowledge transfer and biases embedded in both pretraining and fine-tuning data. In this work, we study multilingual unlearning using the Aya-Expanse 8B model under two settings: (1) data unlearning and (2) concept unlearning. We extend benchmarks for factual knowledge and stereotypes to ten languages through translation: English, French, Arabic, Japanese, Russian, Farsi, Korean, Hindi, Hebrew, and Indonesian. These languages span five language families and a wide range of resource levels. Our experiments show that unlearning in high-resource languages is generally more stable, with asymmetric transfer effects observed between typologically related languages. Furthermore, our analysis of linguistic distances indicates that syntactic similarity is the strongest predictor of cross-lingual unlearning behavior.

Yash More, Prakhar Ganesh, Golnoosh Farnadi

Language models are prone to memorizing their training data, making them vulnerable to extraction attacks. While existing research often examines isolated setups, such as a single model or a fixed prompt, real-world adversaries have a considerably larger attack surface due to access to models across various sizes and checkpoints, and repeated prompting. In this paper, we revisit extraction attacks from an adversarial perspective -- with multi-faceted access to the underlying data. We find significant churn in extraction trends, i.e., even unintuitive changes to the prompt, or targeting smaller models and earlier checkpoints, can extract distinct information. By combining multiple attacks, our adversary doubles ($2 \times$) the extraction risks, persisting even under mitigation strategies like data deduplication. We conclude with four case studies, including detecting pre-training data, copyright violations, extracting personally identifiable information, and attacking closed-source models, showing how our more realistic adversary can outperform existing adversaries in the literature.

Kiarash Mohammadi, Aishwarya Sivaraman, Golnoosh Farnadi

Algorithmic decision making driven by neural networks has become very prominent in applications that directly affect people's quality of life. In this paper, we study the problem of verifying, training, and guaranteeing individual fairness of neural network models. A popular approach for enforcing fairness is to translate a fairness notion into constraints over the parameters of the model. However, such a translation does not always guarantee fair predictions of the trained neural network model. To address this challenge, we develop a counterexample-guided post-processing technique to provably enforce fairness constraints at prediction time. Contrary to prior work that enforces fairness only on points around test or train data, we are able to enforce and guarantee fairness on all points in the input domain. Additionally, we propose an in-processing technique to use fairness as an inductive bias by iteratively incorporating fairness counterexamples in the learning process. We have implemented these techniques in a tool called FETA. Empirical evaluation on real-world datasets indicates that FETA is not only able to guarantee fairness on-the-fly at prediction time but also is able to train accurate models exhibiting a much higher degree of individual fairness.

Ahmed Imtiaz Humayun, Ibtihel Amara, Cristina Vasconcelos et al.

Deep Generative Models are frequently used to learn continuous representations of complex data distributions using a finite number of samples. For any generative model, including pre-trained foundation models with Diffusion or Transformer architectures, generation performance can significantly vary across the learned data manifold. In this paper we study the local geometry of the learned manifold and its relationship to generation outcomes for a wide range of generative models, including DDPM, Diffusion Transformer (DiT), and Stable Diffusion 1.4. Building on the theory of continuous piecewise-linear (CPWL) generators, we characterize the local geometry in terms of three geometric descriptors - scaling ($ψ$), rank ($ν$), and complexity/un-smoothness ($δ$). We provide quantitative and qualitative evidence showing that for a given latent-image pair, the local descriptors are indicative of generation aesthetics, diversity, and memorization by the generative model. Finally, we demonstrate that by training a reward model on the local scaling for Stable Diffusion, we can self-improve both generation aesthetics and diversity using `geometry reward' based guidance during denoising.

Ahmad-Reza Ehyaei, Golnoosh Farnadi, Samira Samadi

Despite the essential need for comprehensive considerations in responsible AI, factors like robustness, fairness, and causality are often studied in isolation. Adversarial perturbation, used to identify vulnerabilities in models, and individual fairness, aiming for equitable treatment of similar individuals, despite initial differences, both depend on metrics to generate comparable input data instances. Previous attempts to define such joint metrics often lack general assumptions about data or structural causal models and were unable to reflect counterfactual proximity. To address this, our paper introduces a causal fair metric formulated based on causal structures encompassing sensitive attributes and protected causal perturbation. To enhance the practicality of our metric, we propose metric learning as a method for metric estimation and deployment in real-world problems in the absence of structural causal models. We also demonstrate the application of our novel metric in classifiers. Empirical evaluation of real-world and synthetic datasets illustrates the effectiveness of our proposed metric in achieving an accurate classifier with fairness, resilience to adversarial perturbations, and a nuanced understanding of causal relationships.

Saber Malekmohammadi, Golnoosh Farnadi

Low-rank adaptation of language models has been proposed to reduce the computational and memory overhead of fine-tuning pre-trained language models. LoRA incorporates trainable low-rank matrices into some parameters of the pre-trained model, called adapters. In this work, we show theoretically that the low-rank adaptation mechanism of LoRA is equivalent to fine-tuning adapters with noisy batch gradients, with the noise variance being a decreasing function of adaptation rank ($r$). Motivated by this understanding, we prove inherent differential privacy for LoRA when adaptation matrices $A_\ell$ are frozen. We show that various factors, e.g., the adaptation rank and batch size, affect the guaranteed privacy level. Our findings provide useful insights into LoRA and uncovers the reason behind the robustness of models fine-tuned with LoRA to privacy attacks.

Rebecca Salganik, Fernando Diaz, Golnoosh Farnadi

Graph neural network (GNN) based methods have saturated the field of recommender systems. The gains of these systems have been significant, showing the advantages of interpreting data through a network structure. However, despite the noticeable benefits of using graph structures in recommendation tasks, this representational form has also bred new challenges which exacerbate the complexity of mitigating algorithmic bias. When GNNs are integrated into downstream tasks, such as recommendation, bias mitigation can become even more difficult. Furthermore, the intractability of applying existing methods of fairness promotion to large, real world datasets places even more serious constraints on mitigation attempts. Our work sets out to fill in this gap by taking an existing method for promoting individual fairness on graphs and extending it to support mini-batch, or sub-sample based, training of a GNN, thus laying the groundwork for applying this method to a downstream recommendation task. We evaluate two popular GNN methods: Graph Convolutional Network (GCN), which trains on the entire graph, and GraphSAGE, which uses probabilistic random walks to create subgraphs for mini-batch training, and assess the effects of sub-sampling on individual fairness. We implement an individual fairness notion called \textit{REDRESS}, proposed by Dong et al., which uses rank optimization to learn individual fair node, or item, embeddings. We empirically show on two real world datasets that GraphSAGE is able to achieve, not just, comparable accuracy, but also, improved fairness as compared with the GCN model. These finding have consequential ramifications to individual fairness promotion, GNNs, and in downstream form, recommender systems, showing that mini-batch training facilitate individual fairness promotion by allowing for local nuance to guide the process of fairness promotion in representation learning.

Edie Pearman, Sophia Osborne, Mira Kandlikar-Bloch et al.

Large language models (LLMs) are increasingly deployed in socially sensitive settings despite substantial documentation that they encode gender biases. Chain-of-Thought (CoT) prompting has been proposed as a bias-mitigation approach. However, existing evaluations primarily focus on changes in LLM benchmark performance, providing limited insight into whether apparent bias reductions reflect meaningful changes in a model's internal mechanisms. In this work, we investigate how CoT prompting affects gender bias in LLMs, combining benchmark-based evaluation with mechanistic interpretability techniques and reasoning chain failure analysis. Our results confirm a stereotypical bias present in LLM outputs across benchmarks, showing that CoT prompting does not consistently reduce the bias gap. Mechanistic analyses reveal that although CoT balances biased behavior in certain attention head clusters, gender bias remains embedded in hidden representations, indicating only superficial mitigation. Inspection of reasoning chains further suggests that these improvements stem from memorization and familiarity with the dataset rather than genuine understanding of bias.

Jesse Thibodeau, Hadi Nekoei, Afaf Taïk et al.

Dynamic, risk-based pricing can systematically exclude vulnerable consumer groups from essential resources such as health insurance and consumer credit. We show that a regulator can realign private incentives with social objectives through a learned, interpretable tax schedule. First, we provide a formal proposition that bounding each firm's \emph{local} demographic gap implicitly bounds the \emph{global} opt-out disparity, motivating firm-level penalties. Building on this insight we introduce \texttt{MarketSim} -- an open-source, scalable simulator of heterogeneous consumers and profit-maximizing firms -- and train a reinforcement learning (RL) social planner (SP) that selects a bracketed fairness-tax while remaining close to a simple linear prior via an $\mathcal{L}_1$ regularizer. The learned policy is thus both transparent and easily interpretable. In two empirically calibrated markets, i.e., U.S. health-insurance and consumer-credit, our planner simultaneously raises demand-fairness by up to $16\%$ relative to unregulated Free Market while outperforming a fixed linear schedule in terms of social welfare without explicit coordination. These results illustrate how AI-assisted regulation can convert a competitive social dilemma into a win-win equilibrium, providing a principled and practical framework for fairness-aware market oversight.

Zhuan Shi, Alireza Dehghanpour Farashah, Rik de Vries et al.

Concept erasure in text-to-image diffusion models seeks to remove undesired concepts while preserving overall generative capability. Localized erasure methods aim to restrict edits to the spatial region occupied by the target concept. However, we observe that suppressing a concept can unintentionally weaken semantically related neighbor concepts, reducing fidelity in fine-grained domains. We propose Neighbor-Aware Localized Concept Erasure (NLCE), a training-free framework designed to better preserve neighboring concepts while removing target concepts. It operates in three stages: (1) a spectrally-weighted embedding modulation that attenuates target concept directions while stabilizing neighbor concept representations, (2) an attention-guided spatial gate that identifies regions exhibiting residual concept activation, and (3) a spatially-gated hard erasure that eliminates remaining traces only where necessary. This neighbor-aware pipeline enables localized concept removal while maintaining the surrounding concept neighborhood structure. Experiments on fine-grained datasets (Oxford Flowers, Stanford Dogs) show that our method effectively removes target concepts while better preserving closely related categories. Additional results on celebrity identity, explicit content and artistic style demonstrate robustness and generalization to broader erasure scenarios.

William St-Arnaud, Margarida Carvalho, Golnoosh Farnadi

Overparameterized fully-connected neural networks have been shown to behave like kernel models when trained with gradient descent, under mild conditions on the width, the learning rate, and the parameter initialization. In the limit of infinitely large widths and small learning rate, the kernel that is obtained allows to represent the output of the learned model with a closed-form solution. This closed-form solution hinges on the invertibility of the limiting kernel, a property that often holds on real-world datasets. In this work, we analyze the sensitivity of large ReLU networks to increasing depths by characterizing the corresponding limiting kernel. Our theoretical results demonstrate that the normalized limiting kernel approaches the matrix of ones. In contrast, they show the corresponding closed-form solution approaches a fixed limit on the sphere. We empirically evaluate the order of magnitude in network depth required to observe this convergent behavior, and we describe the essential properties that enable the generalization of our results to other kernels.

Khaoula Chehbouni, Megha Roshan, Emmanuel Ma et al.

Recent progress in large language models (LLMs) has led to their widespread adoption in various domains. However, these advancements have also introduced additional safety risks and raised concerns regarding their detrimental impact on already marginalized populations. Despite growing mitigation efforts to develop safety safeguards, such as supervised safety-oriented fine-tuning and leveraging safe reinforcement learning from human feedback, multiple concerns regarding the safety and ingrained biases in these models remain. Furthermore, previous work has demonstrated that models optimized for safety often display exaggerated safety behaviors, such as a tendency to refrain from responding to certain requests as a precautionary measure. As such, a clear trade-off between the helpfulness and safety of these models has been documented in the literature. In this paper, we further investigate the effectiveness of safety measures by evaluating models on already mitigated biases. Using the case of Llama 2 as an example, we illustrate how LLMs' safety responses can still encode harmful assumptions. To do so, we create a set of non-toxic prompts, which we then use to evaluate Llama models. Through our new taxonomy of LLMs responses to users, we observe that the safety/helpfulness trade-offs are more pronounced for certain demographic groups which can lead to quality-of-service harms for marginalized populations.

Prakhar Ganesh, Afaf Taik, Golnoosh Farnadi

Algorithmic modeling relies on limited information in data to extrapolate outcomes for unseen scenarios, often embedding an element of arbitrariness in its decisions. A perspective on this arbitrariness that has recently gained interest is multiplicity-the study of arbitrariness across a set of "good models", i.e., those likely to be deployed in practice. In this work, we systemize the literature on multiplicity by: (a) formalizing the terminology around model design choices and their contribution to arbitrariness, (b) expanding the definition of multiplicity to incorporate underrepresented forms beyond just predictions and explanations, (c) clarifying the distinction between multiplicity and other lenses of arbitrariness, i.e., uncertainty and variance, and (d) distilling the benefits and potential risks of multiplicity into overarching trends, situating it within the broader landscape of responsible AI. We conclude by identifying open research questions and highlighting emerging trends in this young but rapidly growing area of research.

Khaoula Chehbouni, Jonathan Colaço Carr, Yash More et al.

In an effort to mitigate the harms of large language models (LLMs), learning from human feedback (LHF) has been used to steer LLMs towards outputs that are intended to be both less harmful and more helpful. Despite the widespread adoption of LHF in practice, the quality of this feedback and its effectiveness as a safety mitigation technique remain unclear. This study addresses these issues by auditing the widely-used Helpful and Harmless (HH) dataset by Anthropic. Our work includes: (1) a thorough investigation of the dataset's content through both manual and automated evaluation; (2) experiments demonstrating the dataset's impact on models' safety; and (3) an analysis of the 100 most influential papers citing this dataset. Through our audit, we showcase how conceptualization failures and quality issues identified in the HH dataset can create additional harms by leading to disparate safety behaviors across demographic groups. Our findings highlight the need for more nuanced, context-sensitive approaches to safety mitigation in LLMs.

Cléa Chataigner, Afaf Taïk, Golnoosh Farnadi

Large language models (LLMs) are increasingly used as alternatives to traditional search engines given their capacity to generate text that resembles human language. However, this shift is concerning, as LLMs often generate hallucinations, misleading or false information that appears highly credible. In this study, we explore the phenomenon of hallucinations across multiple languages in freeform text generation, focusing on what we call multilingual hallucination gaps. These gaps reflect differences in the frequency of hallucinated answers depending on the prompt and language used. To quantify such hallucinations, we used the FactScore metric and extended its framework to a multilingual setting. We conducted experiments using LLMs from the LLaMA, Qwen, and Aya families, generating biographies in 19 languages and comparing the results to Wikipedia pages. Our results reveal variations in hallucination rates, especially between high and low resource languages, raising important questions about LLM multilingual performance and the challenges in evaluating hallucinations in multilingual freeform text generation.

Khaoula Chehbouni, Mohammed Haddou, Jackie Chi Kit Cheung et al.

Evaluating natural language generation (NLG) systems remains a core challenge of natural language processing (NLP), further complicated by the rise of large language models (LLMs) that aims to be general-purpose. Recently, large language models as judges (LLJs) have emerged as a promising alternative to traditional metrics, but their validity remains underexplored. This position paper argues that the current enthusiasm around LLJs may be premature, as their adoption has outpaced rigorous scrutiny of their reliability and validity as evaluators. Drawing on measurement theory from the social sciences, we identify and critically assess four core assumptions underlying the use of LLJs: their ability to act as proxies for human judgment, their capabilities as evaluators, their scalability, and their cost-effectiveness. We examine how each of these assumptions may be challenged by the inherent limitations of LLMs, LLJs, or current practices in NLG evaluation. To ground our analysis, we explore three applications of LLJs: text summarization, data annotation, and safety alignment. Finally, we highlight the need for more responsible evaluation practices in LLJs evaluation, to ensure that their growing role in the field supports, rather than undermines, progress in NLG.

Ibtihel Amara, Ahmed Imtiaz Humayun, Ivana Kajic et al.

Concept erasure techniques have recently gained significant attention for their potential to remove unwanted concepts from text-to-image models. While these methods often demonstrate promising results in controlled settings, their robustness in real-world applications and suitability for deployment remain uncertain. In this work, we (1) identify a critical gap in evaluating sanitized models, particularly in assessing their performance across diverse concept dimensions, and (2) systematically analyze the failure modes of text-to-image models post-erasure. We focus on the unintended consequences of concept removal on non-target concepts across different levels of interconnected relationships including visually similar, binomial, and semantically related concepts. To address this, we introduce EraseBench, a comprehensive benchmark for evaluating post-erasure performance. EraseBench includes over 100 curated concepts, targeted evaluation prompts, and a robust set of metrics to assess both effectiveness and side effects of erasure. Our findings reveal a phenomenon of concept entanglement, where erasure leads to unintended suppression of non-target concepts, causing spillover degradation that manifests as distortions and a decline in generation quality.

Ahmad-Reza Ehyaei, Golnoosh Farnadi, Samira Samadi

In recent years, Wasserstein Distributionally Robust Optimization (DRO) has garnered substantial interest for its efficacy in data-driven decision-making under distributional uncertainty. However, limited research has explored the application of DRO to address individual fairness concerns, particularly when considering causal structures and sensitive attributes in learning problems. To address this gap, we first formulate the DRO problem from causality and individual fairness perspectives. We then present the DRO dual formulation as an efficient tool to convert the DRO problem into a more tractable and computationally efficient form. Next, we characterize the closed form of the approximate worst-case loss quantity as a regularizer, eliminating the max-step in the min-max DRO problem. We further estimate the regularizer in more general cases and explore the relationship between DRO and classical robust optimization. Finally, by removing the assumption of a known structural causal model, we provide finite sample error bounds when designing DRO with empirical distributions and estimated causal structures to ensure efficiency and robust learning.

Florian Carichon, Aditi Khandelwal, Marylou Fauchard et al. · microsoft-research

This position paper states that AI Alignment in Multi-Agent Systems (MAS) should be considered a dynamic and interaction-dependent process that heavily depends on the social environment where agents are deployed, either collaborative, cooperative, or competitive. While AI alignment with human values and preferences remains a core challenge, the growing prevalence of MAS in real-world applications introduces a new dynamic that reshapes how agents pursue goals and interact to accomplish various tasks. As agents engage with one another, they must coordinate to accomplish both individual and collective goals. However, this complex social organization may unintentionally misalign some or all of these agents with human values or user preferences. Drawing on social sciences, we analyze how social structure can deter or shatter group and individual values. Based on these analyses, we call on the AI community to treat human, preferential, and objective alignment as an interdependent concept, rather than isolated problems. Finally, we emphasize the urgent need for simulation environments, benchmarks, and evaluation frameworks that allow researchers to assess alignment in these interactive multi-agent contexts before such dynamics grow too complex to control.

Khaoula Chehbouni, Martine De Cock, Gilles Caporossi et al. · uw

The increased screen time and isolation caused by the COVID-19 pandemic have led to a significant surge in cases of online grooming, which is the use of strategies by predators to lure children into sexual exploitation. Previous efforts to detect grooming in industry and academia have involved accessing and monitoring private conversations through centrally-trained models or sending private conversations to a global server. In this work, we implement a privacy-preserving pipeline for the early detection of sexual predators. We leverage federated learning and differential privacy in order to create safer online spaces for children while respecting their privacy. We investigate various privacy-preserving implementations and discuss their benefits and shortcomings. Our extensive evaluation using real-world data proves that privacy and utility can coexist with only a slight reduction in utility.

Prakhar Ganesh, Usman Gohar, Lu Cheng et al.

With fairness concerns gaining significant attention in Machine Learning (ML), several bias mitigation techniques have been proposed, often compared against each other to find the best method. These benchmarking efforts tend to use a common setup for evaluation under the assumption that providing a uniform environment ensures a fair comparison. However, bias mitigation techniques are sensitive to hyperparameter choices, random seeds, feature selection, etc., meaning that comparison on just one setting can unfairly favour certain algorithms. In this work, we show significant variance in fairness achieved by several algorithms and the influence of the learning pipeline on fairness scores. We highlight that most bias mitigation techniques can achieve comparable performance, given the freedom to perform hyperparameter optimization, suggesting that the choice of the evaluation parameters-rather than the mitigation technique itself-can sometimes create the perceived superiority of one method over another. We hope our work encourages future research on how various choices in the lifecycle of developing an algorithm impact fairness, and trends that guide the selection of appropriate algorithms.

Aly M. Kassem, Zhuan Shi, Negar Rostamzadeh et al.

Large language models (LLMs) are frequently fine-tuned or unlearned to adapt to new tasks or eliminate undesirable behaviors. While existing evaluation methods assess performance after such interventions, there remains no general approach for detecting unintended side effects, such as unlearning biology content degrading performance on chemistry tasks, particularly when these effects are unpredictable or emergent. To address this issue, we introduce MNEME, Model diffiNg for Evaluating Mechanistic Effects, a lightweight framework for identifying these side effects using sparse model diffing. MNEME compares base and fine-tuned models on task-agnostic data (for example, The Pile, LMSYS-Chat-1M) without access to fine-tuning data to isolate behavioral shifts. Applied to five LLMs across three scenarios: WMDP knowledge unlearning, emergent misalignment, and benign fine-tuning, MNEME achieves up to 95 percent accuracy in predicting side effects, aligning with known benchmarks and requiring no custom heuristics. Furthermore, we show that retraining on high-activation samples can partially reverse these effects. Our results demonstrate that sparse probing and diffing offer a scalable and automated lens into fine-tuning-induced model changes, providing practical tools for understanding and managing LLM behavior.

Afaf Taik, Khaoula Chehbouni, Golnoosh Farnadi

Fairness in federated learning has emerged as a rapidly growing area of research, with numerous works proposing formal definitions and algorithmic interventions. Yet, despite this technical progress, fairness in FL is often defined and evaluated in ways that abstract away from the sociotechnical contexts in which these systems are deployed. In this paper, we argue that existing approaches tend to optimize narrow system level metrics, such as performance parity or contribution-based rewards, while overlooking how harms arise throughout the FL lifecycle and how they impact diverse stakeholders. We support this claim through a critical analysis of the literature, based on a systematic annotation of papers for their fairness definitions, design decisions, evaluation practices, and motivating use cases. Our analysis reveals five recurring pitfalls: 1) fairness framed solely through the lens of server client architecture, 2) a mismatch between simulations and motivating use-cases and contexts, 3) definitions that conflate protecting the system with protecting its users, 4) interventions that target isolated stages of the lifecycle while neglecting upstream and downstream effects, 5) and a lack of multi-stakeholder alignment where multiple fairness definitions can be relevant at once. Building on these insights, we propose a harm centered framework that links fairness definitions to concrete risks and stakeholder vulnerabilities. We conclude with recommendations for more holistic, context-aware, and accountable fairness research in FL.

Rohan Sukumaran, Aarash Feizi, Adriana Romero-Sorian et al.

Recent advances in parameter-efficient fine-tuning methods, such as Low Rank Adaptation (LoRA), have gained significant attention for their ability to efficiently adapt large foundational models to various downstream tasks. These methods are appreciated for achieving performance comparable to full fine-tuning on aggregate-level metrics, while significantly reducing computational costs. To systematically address fairness in LLMs previous studies fine-tune on fairness specific data using a larger LoRA rank than typically used. In this paper, we introduce FairLoRA, a novel fairness-specific regularizer for LoRA aimed at reducing performance disparities across data subgroups by minimizing per-class variance in loss. To the best of our knowledge, we are the first to introduce a fairness based finetuning through LoRA. Our results demonstrate that the need for higher ranks to mitigate bias is not universal; it depends on factors such as the pre-trained model, dataset, and task. More importantly, we systematically evaluate FairLoRA across various vision models, including ViT, DiNO, and CLIP, in scenarios involving distribution shifts. We further emphasize the necessity of using multiple fairness metrics to obtain a holistic assessment of fairness, rather than relying solely on the metric optimized during training.

Jesse Thibodeau, Hadi Nekoei, Afaf Taïk et al.

The use of dynamic pricing by profit-maximizing firms gives rise to demand fairness concerns, measured by discrepancies in consumer groups' demand responses to a given pricing strategy. Notably, dynamic pricing may result in buyer distributions unreflective of those of the underlying population, which can be problematic in markets where fair representation is socially desirable. To address this, policy makers might leverage tools such as taxation and subsidy to adapt policy mechanisms dependent upon their social objective. In this paper, we explore the potential for AI methods to assist such intervention strategies. To this end, we design a basic simulated economy, wherein we introduce a dynamic social planner (SP) to generate corporate taxation schedules geared to incentivizing firms towards adopting fair pricing behaviours, and to use the collected tax budget to subsidize consumption among underrepresented groups. To cover a range of possible policy scenarios, we formulate our social planner's learning problem as a multi-armed bandit, a contextual bandit and finally as a full reinforcement learning (RL) problem, evaluating welfare outcomes from each case. To alleviate the difficulty in retaining meaningful tax rates that apply to less frequently occurring brackets, we introduce FairReplayBuffer, which ensures that our RL agent samples experiences uniformly across a discretized fairness space. We find that, upon deploying a learned tax and redistribution policy, social welfare improves on that of the fairness-agnostic baseline, and approaches that of the analytically optimal fairness-aware baseline for the multi-armed and contextual bandit settings, and surpassing it by 13.19% in the full RL setting.

Prakhar Ganesh, Hsiang Hsu, Golnoosh Farnadi

Multiplicity -- the existence of distinct models with comparable performance -- has received growing attention in recent years. While prior work has largely emphasized modelling choices, the critical role of data in shaping multiplicity has been comparatively overlooked. In this work, we introduce a neighbouring datasets framework to examine the most granular case: the impact of a single-data-point difference on multiplicity. Our analysis yields a seemingly counterintuitive finding: neighbouring datasets with greater inter-class distribution overlap exhibit lower multiplicity. This reversal of conventional expectations arises from a shared Rashomon parameter, and we substantiate it with rigorous proofs. Building on this foundation, we extend our framework to two practical domains: active learning and data imputation. For each, we establish natural extensions of the neighbouring datasets perspective, conduct the first systematic study of multiplicity in existing algorithms, and finally, propose novel multiplicity-aware methods, namely, multiplicity-aware data acquisition strategies for active learning and multiplicity-aware data imputation techniques.

Ahmad-Reza Ehyaei, Golnoosh Farnadi, Samira Samadi

Distributionally robust optimization tackles out-of-sample issues like overfitting and distribution shifts by adopting an adversarial approach over a range of possible data distributions, known as the ambiguity set. To balance conservatism and accuracy, these sets must include realistic probability distributions by leveraging information from the nominal distribution. Assuming that nominal distributions arise from a structural causal model with a directed acyclic graph $\mathcal{G}$ and structural equations, previous methods such as adapted and $\mathcal{G}$-causal optimal transport have only utilized causal graph information in designing ambiguity sets. In this work, we propose incorporating structural equations, which include causal graph information, to enhance ambiguity sets, resulting in more realistic distributions. We introduce structural causal optimal transport and its associated ambiguity set, demonstrating their advantages and connections to previous methods. A key benefit of our approach is a relaxed version, where a regularization term replaces the complex causal constraints, enabling an efficient algorithm via difference-of-convex programming to solve structural causal optimal transport. We also show that when structural information is absent and must be estimated, our approach remains effective and provides finite sample guarantees. Lastly, we address the radius of ambiguity sets, illustrating how our method overcomes the curse of dimensionality in optimal transport problems, achieving faster shrinkage with dimension-free order.

Ahmad-Reza Ehyaei, Golnoosh Farnadi, Samira Samadi

Group-fairness metrics (e.g., equalized odds) can vary sharply across resamples and are especially brittle under distribution shift, undermining reliable audits. We propose a Wasserstein distributionally robust framework that certifies worst-case group fairness over a ball of plausible test distributions centered at the empirical law. Our formulation unifies common group fairness notions via a generic conditional-probability functional and defines $\varepsilon$-Wasserstein Distributional Fairness ($\varepsilon$-WDF) as the audit target. Leveraging strong duality, we derive tractable reformulations and an efficient estimator (DRUNE) for $\varepsilon$-WDF. We prove feasibility and consistency and establish finite-sample certification guarantees for auditing fairness, along with quantitative bounds under smoothness and margin conditions. Across standard benchmarks and classifiers, $\varepsilon$-WDF delivers stable fairness assessments under distribution shift, providing a principled basis for auditing and certifying group fairness beyond observational data.

Marylou Fauchard, Florian Carichon, Margarida Carvalho et al.

Recent advances in reasoning with large language models (LLMs) have demonstrated strong performance on complex mathematical tasks, including combinatorial optimization. Techniques such as Chain-of-Thought and In-Context Learning have further enhanced this capability, making LLMs both powerful and accessible tools for a wide range of users, including non-experts. However, applying LLMs to matching problems, which require reasoning under preferential and structural constraints, remains underexplored. To address this gap, we introduce a novel benchmark of 369 instances of the College Admission Problem, a canonical example of a matching problem with preferences, to evaluate LLMs across key dimensions: feasibility, stability, and optimality. We employ this benchmark to assess the performance of several open-weight LLMs. Our results first reveal that while LLMs can satisfy certain constraints, they struggle to meet all evaluation criteria consistently. They also show that reasoning LLMs, like QwQ and GPT-oss, significantly outperform traditional models such as Llama, Qwen or Mistral, defined here as models used without any dedicated reasoning mechanisms. Moreover, we observed that LLMs reacted differently to the various prompting strategies tested, which include Chain-of-Thought, In-Context Learning and role-based prompting, with no prompt consistently offering the best performance. Finally, we report the performances from iterative prompting with auto-generated feedback and show that they are not monotonic; they can peak early and then significantly decline in later attempts. Overall, this work offers a new perspective on model reasoning performance and the effectiveness of prompting strategies in combinatorial optimization problems with preferential constraints.

Cléa Chataigner, Rebecca Ma, Prakhar Ganesh et al.

Large language models (LLMs) are highly sensitive to subtle changes in prompt phrasing, posing challenges for reliable auditing. Prior methods often apply unconstrained prompt paraphrasing, which risk missing linguistic and demographic factors that shape authentic user interactions. We introduce AUGMENT (Automated User-Grounded Modeling and Evaluation of Natural Language Transformations), a framework for generating controlled paraphrases, grounded in user behaviors. AUGMENT leverages linguistically informed rules and enforces quality through checks on instruction adherence, semantic similarity, and realism, ensuring paraphrases are both reliable and meaningful for auditing. Through case studies on the BBQ and MMLU datasets, we show that controlled paraphrases uncover systematic weaknesses that remain obscured under unconstrained variation. These results highlight the value of the AUGMENT framework for reliable auditing.

Florian Carichon, Romain Rampa, Golnoosh Farnadi

Novelty modeling and detection is a core topic in Natural Language Processing (NLP), central to numerous tasks such as recommender systems and automatic summarization. It involves identifying pieces of text that deviate in some way from previously known information. However, novelty is also a crucial determinant of the unique perception of relevance and quality of an experience, as it rests upon each individual's understanding of the world. Social factors, particularly cultural background, profoundly influence perceptions of novelty and innovation. Cultural novelty arises from differences in salience and novelty as shaped by the distance between distinct communities. While cultural diversity has garnered increasing attention in artificial intelligence (AI), the lack of robust metrics for quantifying cultural novelty hinders a deeper understanding of these divergences. This gap limits quantifying and understanding cultural differences within computational frameworks. To address this, we propose an interdisciplinary framework that integrates knowledge from sociology and management. Central to our approach is GlobalFusion, a novel dataset comprising 500 dishes and approximately 100,000 cooking recipes capturing cultural adaptation from over 150 countries. By introducing a set of Jensen-Shannon Divergence metrics for novelty, we leverage this dataset to analyze textual divergences when recipes from one community are modified by another with a different cultural background. The results reveal significant correlations between our cultural novelty metrics and established cultural measures based on linguistic, religious, and geographical distances. Our findings highlight the potential of our framework to advance the understanding and measurement of cultural diversity in AI.

Armin Moradi, Nicola Neophytou, Florian Carichon et al.

Popularity bias in recommender systems can increase cultural overrepresentation by favoring norms from dominant cultures and marginalizing underrepresented groups. This issue is critical for platforms offering cultural products, as they influence consumption patterns and human perceptions. In this work, we address popularity bias by identifying demographic biases within prototype-based matrix factorization methods. Using the country of origin as a proxy for cultural identity, we link this demographic attribute to popularity bias by refining the embedding space learning process. First, we propose filtering out irrelevant prototypes to improve representativity. Second, we introduce a regularization technique to enforce a uniform distribution of prototypes within the embedding space. Across four datasets, our results demonstrate a 27\% reduction in the average rank of long-tail items and a 2\% reduction in the average rank of items from underrepresented countries. Additionally, our model achieves a 2\% improvement in HitRatio@10 compared to the state-of-the-art, highlighting that fairness is enhanced without compromising recommendation quality. Moreover, the distribution of prototypes leads to more inclusive explanations by better aligning items with diverse prototypes.

Shahrad Mohammadzadeh, Juan David Guerra, Marco Bonizzato et al.

As large language models (LLMs) become increasingly prevalent, concerns about their reliability, particularly due to hallucinations - factually inaccurate or irrelevant outputs - have grown. Our research investigates the relationship between the uncertainty in training dynamics and the emergence of hallucinations. Using models from the Pythia suite and several hallucination detection metrics, we analyze hallucination trends and identify significant variance during training. To address this, we propose Sensitivity Dropout (SenD), a novel training protocol designed to reduce hallucination variance during training by deterministically dropping embedding indices with significant variability. In addition, we develop an unsupervised hallucination detection metric, Efficient EigenScore (EES), which approximates the traditional EigenScore in 2x speed. This metric is integrated into our training protocol, allowing SenD to be both computationally scalable and effective at reducing hallucination variance. SenD improves test-time reliability of Pythia and Meta's Llama models by up to 17% and enhances factual accuracy in Wikipedia, Medical, Legal, and Coding domains without affecting downstream task performance.

Golnoosh Farnadi, Mohammad Havaei, Negar Rostamzadeh

The rise of foundation models holds immense promise for advancing AI, but this progress may amplify existing risks and inequalities, leaving marginalized communities behind. In this position paper, we discuss that disparities towards marginalized communities - performance, representation, privacy, robustness, interpretability and safety - are not isolated concerns but rather interconnected elements of a cascading disparity phenomenon. We contrast foundation models with traditional models and highlight the potential for exacerbated disparity against marginalized communities. Moreover, we emphasize the unique threat of cascading impacts in foundation models, where interconnected disparities can trigger long-lasting negative consequences, specifically to the people on the margin. We define marginalized communities within the machine learning context and explore the multifaceted nature of disparities. We analyze the sources of these disparities, tracing them from data creation, training and deployment procedures to highlight the complex technical and socio-technical landscape. To mitigate the pressing crisis, we conclude with a set of calls to action to mitigate disparity at its source.

Sikha Pentyala, David Melanson, Martine De Cock et al.

Machine learning (ML) has become prominent in applications that directly affect people's quality of life, including in healthcare, justice, and finance. ML models have been found to exhibit discrimination based on sensitive attributes such as gender, race, or disability. Assessing if an ML model is free of bias remains challenging to date, and by definition has to be done with sensitive user characteristics that are subject of anti-discrimination and data protection law. Existing libraries for fairness auditing of ML models offer no mechanism to protect the privacy of the audit data. We present PrivFair, a library for privacy-preserving fairness audits of ML models. Through the use of Secure Multiparty Computation (MPC), PrivFair protects the confidentiality of the model under audit and the sensitive data used for the audit, hence it supports scenarios in which a proprietary classifier owned by a company is audited using sensitive audit data from an external investigator. We demonstrate the use of PrivFair for group fairness auditing with tabular data or image data, without requiring the investigator to disclose their data to anyone in an unencrypted manner, or the model owner to reveal their model parameters to anyone in plaintext.

Aishwarya Sivaraman, Golnoosh Farnadi, Todd Millstein et al.

The widespread adoption of deep learning is often attributed to its automatic feature construction with minimal inductive bias. However, in many real-world tasks, the learned function is intended to satisfy domain-specific constraints. We focus on monotonicity constraints, which are common and require that the function's output increases with increasing values of specific input features. We develop a counterexample-guided technique to provably enforce monotonicity constraints at prediction time. Additionally, we propose a technique to use monotonicity as an inductive bias for deep learning. It works by iteratively incorporating monotonicity counterexamples in the learning process. Contrary to prior work in monotonic learning, we target general ReLU neural networks and do not further restrict the hypothesis space. We have implemented these techniques in a tool called COMET. Experiments on real-world datasets demonstrate that our approach achieves state-of-the-art results compared to existing monotonic learners, and can improve the model quality compared to those that were trained without taking monotonicity constraints into account.

Golnoosh Farnadi, Lise Getoor, Marie-Francine Moens et al.

A variety of approaches have been proposed to automatically infer the profiles of users from their digital footprint in social media. Most of the proposed approaches focus on mining a single type of information, while ignoring other sources of available user-generated content (UGC). In this paper, we propose a mechanism to infer a variety of user characteristics, such as, age, gender and personality traits, which can then be compiled into a user profile. To this end, we model social media users by incorporating and reasoning over multiple sources of UGC as well as social relations. Our model is based on a statistical relational learning framework using Hinge-loss Markov Random Fields (HL-MRFs), a class of probabilistic graphical models that can be defined using a set of first-order logical rules. We validate our approach on data from Facebook with more than 5k users and almost 725k relations. We show how HL-MRFs can be used to develop a generic and extensible user profiling framework by leveraging textual, visual, and relational content in the form of status updates, profile pictures and Facebook page likes. Our experimental results demonstrate that our proposed model successfully incorporates multiple sources of information and outperforms competing methods that use only one source of information or an ensemble method across the different sources for modeling of users in social media.

Behrouz Babaki, Golnoosh Farnadi, Gilles Pesant

Factored stochastic constraint programming (FSCP) is a formalism to represent multi-stage decision making problems under uncertainty. FSCP models support factorized probabilistic models and involve constraints over decision and random variables. These models have many applications in real-world problems. However, solving these problems requires evaluating the best course of action for each possible outcome of the random variables and hence is computationally challenging. FSCP problems often involve repeated subproblems which ideally should be solved once. In this paper we show how identifying and exploiting these identical subproblems can simplify solving them and leads to a compact representation of the solution. We compile an And-Or search tree to a compact decision diagram. Preliminary experiments show that our proposed method significantly improves the search efficiency by reducing the size of the problem and outperforms the existing methods.

YooJung Choi, Golnoosh Farnadi, Behrouz Babaki et al.

As machine learning is increasingly used to make real-world decisions, recent research efforts aim to define and ensure fairness in algorithmic decision making. Existing methods often assume a fixed set of observable features to define individuals, but lack a discussion of certain features not being observed at test time. In this paper, we study fairness of naive Bayes classifiers, which allow partial observations. In particular, we introduce the notion of a discrimination pattern, which refers to an individual receiving different classifications depending on whether some sensitive attributes were observed. Then a model is considered fair if it has no such pattern. We propose an algorithm to discover and mine for discrimination patterns in a naive Bayes classifier, and show how to learn maximum likelihood parameters subject to these fairness constraints. Our approach iteratively discovers and eliminates discrimination patterns until a fair model is learned. An empirical evaluation on three real-world datasets demonstrates that we can remove exponentially many discrimination patterns by only adding a small fraction of them as constraints.

Golnoosh Farnadi, Pigi Kouki, Spencer K. Thompson et al.

Recommender systems are used in variety of domains affecting people's lives. This has raised concerns about possible biases and discrimination that such systems might exacerbate. There are two primary kinds of biases inherent in recommender systems: observation bias and bias stemming from imbalanced data. Observation bias exists due to a feedback loop which causes the model to learn to only predict recommendations similar to previous ones. Imbalance in data occurs when systematic societal, historical, or other ambient bias is present in the data. In this paper, we address both biases by proposing a hybrid fairness-aware recommender system. Our model provides efficient and accurate recommendations by incorporating multiple user-user and item-item similarity measures, content, and demographic information, while addressing recommendation biases. We implement our model using a powerful and expressive probabilistic programming language called probabilistic soft logic. We experimentally evaluate our approach on a popular movie recommendation dataset, showing that our proposed model can provide more accurate and fairer recommendations, compared to a state-of-the art fair recommender system.

Sisi Wang, Wing-Sea Poon, Golnoosh Farnadi et al.

User profiling from user generated content (UGC) is a common practice that supports the business models of many social media companies. Existing systems require that the UGC is fully exposed to the module that constructs the user profiles. In this paper we show that it is possible to build user profiles without ever accessing the user's original data, and without exposing the trained machine learning models for user profiling -- which are the intellectual property of the company -- to the users of the social media site. We present VirtualIdentity, an application that uses secure multi-party cryptographic protocols to detect the age, gender and personality traits of users by classifying their user-generated text and personal pictures with trained support vector machine models in a privacy-preserving manner.

Varun Embar, Dhanya Sridhar, Golnoosh Farnadi et al.

Statistical relational frameworks such as Markov logic networks and probabilistic soft logic (PSL) encode model structure with weighted first-order logical clauses. Learning these clauses from data is referred to as structure learning. Structure learning alleviates the manual cost of specifying models. However, this benefit comes with high computational costs; structure learning typically requires an expensive search over the space of clauses which involves repeated optimization of clause weights. In this paper, we propose the first two approaches to structure learning for PSL. We introduce a greedy search-based algorithm and a novel optimization method that trade-off scalability and approximations to the structure learning problem in varying ways. The highly scalable optimization method combines data-driven generation of clauses with a piecewise pseudolikelihood (PPLL) objective that learns model structure by optimizing clause weights only once. We compare both methods across five real-world tasks, showing that PPLL achieves an order of magnitude runtime speedup and AUC gains up to 15% over greedy search.