Scientific Publications of the University of Toulouse II Le Mirail
Not a member yet
92205 research outputs found
Sort by
PatientDx: Merging Large Language Models for Protecting Data-Privacy in Healthcare
No full textInternational audienceFine-tuning of Large Language Models (LLMs) has become the default practice for improving model performance on a given task. However, performance improvement comes at the cost of training on vast amounts of annotated data which could be sensitive leading to significant data privacy concerns. In particular, the healthcare domain is one of the most sensitive domains exposed to data privacy issues. In this paper, we present PatientDx, a framework of model merging that allows the design of effective LLMs for health-predictive tasks without requiring fine-tuning nor adaptation on patient data. Our proposal is based on recently proposed techniques known as merging of LLMs and aims to optimize a building block merging strategy. PatientDx uses a pivotal model adapted to numerical reasoning and tunes hyperparameters on examples based on a performance metric but without training of the LLM on these data. Experiments using the mortality tasks of the MIMIC-IV dataset show improvements up to 7% in terms of AUROC when compared to initial models. Additionally, we confirm that when compared to fine-tuned models, our proposal is less prone to data leak problems without hurting performance. Finally, we qualitatively show the capabilities of our proposal through a case study. Our best model is publicly available at https://huggingface.co/ Jgmorenof/mistral_merged_0_4.</div
Anti-immigration conspiracy beliefs are associated with endorsement of conventional and violent actions opposing immigration and attitudes towards democracy across 21 countries
No full textInternational audienceDespite widespread speculation that conspiracy beliefs foster anti-democratic outcomes, the empirical picture is inconsistent. To clarify this literature, we examine the relationships that conspiracy beliefs have with commitment to reactionary action and criticism of democracy, focusing on a global issue: immigration. We expected that people who believe that their government uses immigration to diversify the population against citizens' wishes (anti-migration conspiracy beliefs) would be more committed to conventional and violent action to oppose immigration, and more critical of democracy. However, societal-level factorseconomic performance and democratic functioningwere expected to influence (strengthen, weaken) these links. As hypothesized, multi-level analyses (N = 4353) from 21 countries revealed that economic prosperity attenuated the positive link between anti-migration conspiracy beliefs and commitment to reactionary action. Paradoxically, more democratic societies evidenced stronger links between conspiracy beliefs and conventional (but not violent) action to oppose immigration. Thus, more democratic societies appear to invite conventional forms of action to oppose immigration which may, in turn, weaken democratic norms of inclusion. Results highlight the interplay of individual-and societal-level factors underlying illiberal movements.</div
Levé topographique par GPS différentiels d'un site de fouilles à Ste-Colome (64) Acquisition, traitements et restitutions
No full text
Navigating Partial Automation in Firefighting with Drones: Trust, Take-Over, and Human-Drone Teaming
No full textInternational audienceIn the past years, the use of drones has been increasingly introduced to firefighting operations. Drawing on concrete examples from interviews with Thai firefighting professionals and recent field trials, as well as prior research, this paper examines the challenges of integrating (partially) autonomous, AI-enhanced drones into firefighting operations. Our findings reveal that, despite the promise of automation, on-field operators still prefer communication via a dedicated drone pilot-a preference driven by unresolved trust issues and concerns over information overload. We discuss challenges such as trust in automation and adaptive take-over. These inform our proposals for design recommendations on adaptive communication, transparent take-over mechanisms, trust calibration, physical handover and mapping of multiple data sources in human-drone teaming for firefighting
An asymptotic preserving scheme for the quantum Liouville-BGK equation
No full textWe are interested in this work in the numerical resolution of the Quantum Liouville-BGK equation, which arises in the derivation of quantum hydrodynamical models from first principles. Such models are often obtained in some asymptotic limits, for instance a diffusion or a fluid limit, and as a consequence the original Liouville equation contains small parameters. A standard method such as a split-step algorithm is then accurate provided the time step is sufficiently small compared to the asymptotic parameter, which is a severe limitation. In the case of the diffusion limit, we propose a numerical method that is accurate for time steps independent of the small parameter, and which captures well both the microscopic dynamics and the diffusion limit. Our approach is substantiated by an informal theoretical error analysis
Ordonner, raconter, remémorer. Quelques remarques sur les chapiteaux de la crypte Saint-Girons d’Hagetmau (Landes)
No full textInternational audienc
Simplicity of the automorphism group of fields with operators
No full textWe adapt a proof of Lascar in order to show the simplicity of the group of automorphisms fixing pointwise all non-generic elements for a class of uncountable models of suitable theories of fields with operators
Visual Assistance in VR-based Robot Control: Towards a Reproducible Evaluation Scenario
No full textNational audienceImmersive technologies enhance industrial applications by creating virtual representations of manufacturing environments and providing visual assistance for tasks. Most studies in the literature are based on ad hoc scenarios, making comparison across studies challenging. In this research, 99 participants controlled a remote industrial robot using a VR headset in a reproducible maintenance task. They performed the task under three conditions: without assistance (control), text-based assistance, or attentional cueing (highlighting objects). We manipulated task difficulty (easy versus difficult) and assessed performance based on completion time and failure rate. Results showed visual assistance significantly shortened task completion time, with findings discussed