HAL Paris Dauphine-PSL
Not a member yet
21932 research outputs found
Sort by
Governing prediction: A material political economy of the computational toxicology market
No full textInternational audienceOver the past two decades, prediction has emerged as a service for the chemical industry, with computational tools helping to assess the properties and register chemical substances. This article examines this digitalization of regulatory risk assessment and the specialized industry and tools that took form in the process. In this process of digitalization, prediction strictly speakingestimating the toxicity of molecules using statistical models without prior experimental knowledge of those moleculeshas given way to a different digital practice called 'read-across', a computer-aided extrapolation of toxicological information across groups of known chemicals. A consortium of public actors including the Organisation for Economic Co-operation and Development (OECD) and a series of public scientific and regulatory bodies have developed an open-use tool, the OECD QSAR Toolbox, that has taken the lead over commercially developed tools of prediction. This particular trajectory of digitalization and the resulting market morphology call for more detailed analysis of the dynamics of digital capitalism and the transformation of data political economies. Drawing from a material political economy perspective, we argue that the read-across approach and the OECD toolbox won thanks to the suite of political interventions that made digital analysis doable for the vast majority of small companies subjected to market entry regulation of chemicals, while decisively evaluating and limiting the credibility of commercial, model-based predictions. This result adds to the discussion on data capitalism, highlighting the public and political aspects inherent in the design of digital industries and prediction products
Beyond Politus, Ordinary Democracy in Organization
No full textInternational audienceThiis article is a report on the 15th Organizations, Artifacts & Practices (OAP) workshop held at the London School of Economics. It looks back at three presences and one absence in our discussions on democracy in organizations. More than ever, we need to go beyond the simple posture of politus in democratic conversations to make di'erences productive. And this is a never-ending task
Improving Diversity in Language Models: When Temperature Fails, Change the Loss
No full textInternational audienceIncreasing diversity in language models is a challenging yet essential objective. A common approach is to raise the decoding temperature. In this work, we investigate this approach through a simplistic yet common case to provide insights into why decreasing temperature can improve quality (Precision), while increasing it often fails to boost coverage (Recall). Our analysis reveals that for a model to be effectively tunable through temperature adjustments, it must be trained toward coverage. To address this, we propose rethinking loss functions in language models by leveraging the Precision-Recall framework. Our results demonstrate that this approach achieves a substantially better trade-off between Precision and Recall than merely combining negative log-likelihood training with temperature scaling. These findings offer a pathway toward more versatile and robust language modeling techniques
Perspectives des sciences sociales pour comprendre et sortir de l’antibiodépendance
No full textPPR AntibiorésistanceIt is now common knowledge that antimicrobial resistance (AMR) is a social problem that must not only be studied as a biomedical challenge, but for which other factors – i.e. sociological, economic, political and cultural determinants – must be taken into account. Most research into social factors has been conducted from a social psychological and behaviourist perspective, but it tends to overlook the structural and organisational determinants that provide the framework for individual behaviour. It is precisely these structural determinants (public policies, markets, professions, technologies, etc.) that the teams working on the Digital Observatory of Antibiotic Resistance (DOSA) and the STATIC project, funded by the Antibiotic Resistance RPP, are looking at.Il est désormais de notoriété publique que l’antibiorésistance (AMR) est un problème social qui ne doit pas seulement être étudié comme un défi biomédical mais pour lequel d’autres facteurs – c’est-à-dire des déterminants sociologiques, économiques, politiques, culturels – doivent être pris en compte. La plupart des recherches sur les facteurs sociaux ont été menées dans une perspective de psychologie sociale et de comportementaliste mais ils ont tendance à négliger les déterminants structurels et organisationnels qui fournissent le cadre du comportement individuel. Ce sont précisément ces déterminants structurels (politiques publiques, marchés, professions, technologies, etc.) auxquels s’intéressent les équipes de l’Observatoire numérique de l’Antibiorésistance (DOSA) et du projet STATIC, financés par le PPR Antibiorésistance
fiabilité des modèles de langage génératifs : efficacité, fidélité et couverture
No full textThis thesis investigates how to make generative language models more reliable by addressing two key challenges: precision in conditional generation and diversity in outputs. We study how architectural choices and training methods influence these aspects and propose targeted solutions that improve reliability across different dimensions of text generation.We begin by formalizing text generation with tools from probabilistic modeling, reviewing the training objectives and architectural innovations that led to current large language models. This part provides a unifying perspective for analyzing both conditional generation and distributional coverage.Building on this, we first examine long-context summarization, where existing transformer architectures face computational limitations. We propose an encoder-decoder architecture based on state-space models, which reduces memory requirements and enables processing of far longer sequences while maintaining competitive performance.We then turn to training methods for faithfulness, investigating how models can be prevented from introducing unsupported information. To this end, we introduce Scope, a self-supervised framework that generates unfaithful samples and applies preference-based training to downweight spurious patterns. Scope improves alignment with the input without requiring human annotations, yielding more faithful outputs across multiple evaluation settings.Finally, we study distributional coverage, balancing output quality with diversity. We adapt the Precision--Recall framework, originally proposed for image generation, to text generation, introducing new metrics that reveal systematic trade-offs introduced by fine-tuning. We also show that steering models requires more than temperature adjustments at inference, and propose loss functions grounded in Precision--Recall to better control this trade-off.Cette thèse étudie comment rendre les modèles de langage génératif plus fiables en abordant deux défis majeurs : la précision dans la génération conditionnelle et la diversité des sorties. Nous analysons l’influence des choix architecturaux et des méthodes d’entraînement sur ces aspects et proposons des solutions ciblées qui améliorent la fiabilité selon différentes dimensions de la génération de texte. Nous commençons par formaliser la génération de texte à l’aide d’outils issus de la modélisation probabiliste, en passant en revue les objectifs d’entraînement et les innovations architecturales qui ont conduit aux modèles de langage de grande taille actuels. Cette partie fournit une perspective unificatrice pour analyser à la fois la génération conditionnelle et la couverture distributionnelle. Dans ce cadre, nous examinons d’abord la synthèse de textes longs, un domaine où les architectures de type transformeur rencontrent des limites computationnelles. Nous proposons une architecture encodeur-décodeur basée sur les modèles d’espace d’états, qui réduit les besoins en mémoire et permet de traiter des séquences beaucoup plus longues tout en maintenant des performances compétitives. Nous nous tournons ensuite vers les méthodes d’entraînement pour la fidélité, en étudiant comment empêcher les modèles d’introduire des informations non justifiées. À cette fin, nous introduisons Scope, un cadre auto-supervisé qui génère des exemples infidèles et applique un entraînement par préférences pour atténuer les motifs trompeurs. Scope améliore l’alignement avec l’entrée sans nécessiter d’annotations humaines, produisant ainsi des sorties plus fidèles dans divers contextes d’évaluation. Enfin, nous étudions la couverture distributionnelle, en cherchant à équilibrer qualité et diversité des sorties. Nous adaptons le cadre Précision-Rappel, initialement proposé pour la génération d’images, à la génération de texte, en introduisant de nouvelles métriques qui révèlent les compromis systématiques induits par l’adaptation fine. Nous montrons également que le pilotage des modèles requiert davantage que de simples ajustements de température lors de l’inférence, et proposons des fonctions de perte fondées sur Précision-Rappel pour mieux contrôler ce compromis
The hindered development of continuing education in French universities. A paradoxical approach
No full textInternational audienc
Climate Shocks and U.S. Bank Stability
No full textThis paper investigates the effects of systemic climate variability on U.S. banking stability using the El Niño-Southern Oscillation (ENSO) as a quasi-natural experiment. In contrast to studies focusing on rare, localized natural disasters, we examine how persistent and spatially heterogeneous ENSO-induced climate anomalies-especially those associated with the often-overlooked La Niña phase-affect banks across the continental United States. ENSO is the most influential source of interannual climate variation on Earth and provides a compelling setting to study the transmission of exogenous physical risks to the financial sector. We construct a 30-year quarterly panel of over 800,000 bank-quarter observations (1994-2023), combining detailed financial data with geolocated branch networks and high-resolution teleconnection estimates of local temperature anomalies. Our empirical strategy combines three key elements: a regime-based climate shock identification grounded in recent climate science, a granular spatial matching of institutions to localized exposure, and a dynamic panel framework based on local projections. Our results show that strong La Niña shocks reduce the distance to default by roughly 20%, with effects peaking between 7 and 11 quarters after the shock. These disruptions operate primarily through rising credit risk, lower profitability, and weaker solvency-particularly in banks with large real estate exposure, broad but climate-sensitive geographic footprints, and sizable balance sheets. These findings underscore the need for prudential regulation to incorporate granular, forward-looking metrics of physical climate risk, especially as ongoing climate change is expected to increase the frequency and intensity of ENSO events
Colouring Complete Multipartite and Kneser-Type Digraphs
No full textInternational audienceThe dichromatic number of a digraph is the smallest such that can be partitioned into acyclic subdigraphs, and the dichromatic number of an undirected graph is the maximum dichromatic number over all its orientations. Extending a well-known result of Lovász, we show that the dichromatic number of the Kneser graph is and that the dichromatic number of the Borsuk graph is if is large enough. We then study the list version of the dichromatic number. We show that, for any \varepsilon>0 and , the list dichromatic number of is . This extends a recent result of Bulankina and Kupavskii on the list chromatic number of , where the same behaviour was observed. We also show that for any \rho>3, and , the list dichromatic number of the complete -partite graph with vertices in each part is , extending a classical result of Alon. Finally, we give a directed analogue of Sabidussi's theorem on the chromatic number of graph products