R3AI Research Clusters 2027–2030: projects selected for the second phase of evaluation

La transition vers une économie décarbonée exige une prise de décision stratégique sur les matériaux et procédés à prioriser. Toutefois, ces décisions reposent sur l’intégration de dimensions scientifiques, techno-économiques, sociales et (géo)politiques, incluant l’éthique, l’opinion publique et la justice sociale, qui sont complexes, interreliées, et qui évoluent rapidement. Les données pour guider cette prise de décision sont disponibles, mais souvent de manière morcelée, ce qui limite la capacité des approches actuelles à soutenir une prise de décision éclairée. Des initiatives comme la Feuille de route Ambition 2035 de PRIMA Québec ohrent des portraits de l’écosystème utiles, mais demeurent statiques, sectoriels et ne tiennent pas compte des dynamiques sociales, (géo)politiques ou des avancées scientifiques en temps réel. Le Regroupement développera une plateforme innovante fondée sur une fédération de jumeaux numériques, associés à leur jumeau physique, orchestrés par l’IA agentique. Chaque jumeau simulera une dimension clé (contexte politique et (géo)politique, expertise scientifique, capacités industrielles et contexte socio-économique) afin d’évaluer dihérents scénarios et d’identifier les matériaux et procédés les plus stratégiques pour atteindre les objectifs de décarbonation et de transition énergétique. Pour débuter, la plateforme se concentrera sur les matériaux polymères durables et les procédés de fabrication additive associés. Au-delà de développer des outils de l’IA pour l’accélération de la découverte de nouveaux matériaux, le Regroupement vise à développer un outil IA capable de comprendre l’écosystème des matériaux dans toute sa complexité, à mettre en évidence des relations jusque-là difficiles à établir et à soutenir une prospective stratégique fondée sur des données probantes. Une interface de visualisation interactive facilitera l’utilisation de la plateforme pour toutes parties prenantes, notamment en fournissant des données probantes aux décideurs et en permettant aux citoyens de mieux saisir les enjeux politiques et sociétaux de l’écosystème des matériaux. Le succès de ce Regroupement repose sur une collaboration interdisciplinaire entre les experts de l’IA, des matériaux, des politiques publiques et des sciences sociales, où l’expertise humaine complète les capacités de l’IA pour guider la prise de décisions stratégiques.

Principal investigator

• Audrey Laventure – Université de Montréal

Co-Principal Investigators

• Delphine Bouilly – Université de Montréal
• Bang Liu – Université de Montréal
• Christine Rothmayr – Université de Montréal

Co-investigators

• Aram Amassian – Polytechnique Montréal
• Evelyne Brie – Université de Montréal
• Laurent Charlin – HEC Montréal
• Mickaël Dollé – Université de Montréal
• Frédéric Mérand – Université de Montréal
• Houari Sahraoui – Université de Montréal
• Carlos Silva – Université de Montréal

Artificial Intelligence is transitioning from passive text processing to active, world-grounded reasoning. However, current frontier AI systems remain fundamentally constrained: they hallucinate facts, lack deep environmental grounding, and fail to guarantee safety when executing complex multi-step actions. This research initiative introduces next-generation Language-World Models, which are AI architectures designed to unify linguistic reasoning, visual-spatial perception, and actionable real-world environments within a provably safe framework. Bringing together leading researchers across Mila (Université de Montréal, Polytechnique Montréal, HEC Montréal, UQAM, McGill) and international collaborator institutions (Princeton, UC Davis, CNRS Paris-Saclay, Korea University), this project bridges natural language processing, cognitive science, machine learning, and AI safety. The project advances three core scientific pillars:

World-Grounded Reasoning: Moving beyond statistical word prediction by grounding language models in multimodal latent representation spaces.
Provable Safety and Alignment: Developing rigorous algorithmic guardrails, interpretable reasoning structures, and safety guarantees before actions are executed in real-world contexts.
Adaptive Multi-Agent Intelligence: Enabling collaborative AI systems capable of continuous learning, transparent decision-making, and robust interaction in complex environments.
By integrating insights from cognitive development, natural language processing, machine learning, and safety-by-design principles, this initiative addresses the critical reliability bottlenecks of modern AI. The resulting methodologies will yield transformative, trustworthy AI applications for science, medicine, human-robot collaboration, and complex decision support, establishing a new standard for safe, intelligent systems that reason reliably in real-world environments.

L’intelligence artificielle (IA) transforme la société, mais son expansion rapide risque aussi de créer un écart croissant entre la puissance de l’IA et la capacité réelle des apprenant·e·s, créateur·trice·s, grand public, formateur·rice·s, et décideur·se·s à comprendre, accepter et adopter ces changements d’une façon éthique, responsable et informée quant à ses effets. L’IA pourrait ainsi accentuer les inégalités en avantageant les publics, organisations et milieux déjà bien outillés au détriment d’autres qui resteraient à l’écart faute d’accès, de confiance ou de formation pour développer leurs capacités. La question est donc à la fois technologique, sociale, éthique, éducative et pédagogique. Notre Regroupement s’attaque à cet enjeu en bâtissant un écosystème éducatif qui soutient la maîtrise de l’IA dans différents milieux d’apprentissage (p. ex., écoles, cégeps, universités, musées), pour incorporer à la mission éducative d’IVADO l’expertise des sciences de l’éducation et de la mobilisation des connaissances, et structurer les démarches de formation et d’accompagnement des milieux éducatifs et d’une variété de professionnel·le·s (enseignant·e·s, orthopédagogues, conseiller·ère·s pédagogiques, cadres). Nous aurons ainsi un impact sur l’ensemble du milieu éducatif, avec le soutien de deux ministères et de partenaires en éducation.

Nos activités s’organisent en trois axes :
1. Littératie de l’IA, développement des compétences et mobilisation des connaissances
2. Créativité, conception, collaboration et apprentissage avec l’IA
3. Politiques éducatives, éthique, agentivité et IA responsable pour le futur de l’éducation En réunissant sciences de l’éducation, sciences informatiques, arts et création, et technologies, nous anticipons des retombées importantes, incluant des programmes de formation inclusifs et adaptés, la formation des personnes enseignantes et le développement de politiques éducatives pertinentes et efficaces accentuant la souveraineté numérique du Canada.

As the Healthcare system cluster of R3AI IVADO since 2023, our research program moves beyond AI prototypes to real-world responsible integration, thus becoming a collective lever of transformation for our universal healthcare system. Our program comprises four ambitious and strategic research objectives, each led by a working group of interdisciplinary experts to bridge fundamental AI research with systemic integration. The governance team is formed of established and emerging leaders from diverse disciplines, and is supported by a full-time core team of technical and deployment leads. This structure ensures scientific execution remains aligned with clinical needs, organizational realities, and R3AI vision: Robustness: We are developing the infrastructure to transition models to organizational deployment, exploring solutions to safely mobilize national population data, ensuring models are validated against diverse data streams and contextually evaluated across the Quebec universal healthcare system. Reasoning: We are replacing opaque AI systems with agentic orchestration. An AI orchestrator dynamically links patient data to our community bank of generic and specialized models, creating understandable workflows for users. This provides the explainability required to build trust and support complex decision-making. Responsible: We are building sovereign AI foundations by developing the expertise to run models locally and by co-creating an Implementation Toolkit guiding responsible implementation, with adequate legal, ethical and organizational guardrails and practices, leading to optimal AI deployment. This toolkit comprise an evaluation framework to assess the added value of AI based on patients, caregivers, managers, clinicians, and the systems’ needs, including metrics of fairness, sustainability and health equity. Our team will build lasting organizational capacities to ensure that the benefits of safe, accessible and high-value AI are equitably distributed within our universal healthcare system.

Principal investigator

• Christian Gagné – Université Laval

Co-Principal Investigators

• Nadia Lahrichi – Polytechnique Montréal
• Aude Motulsky – Université de Montréal
• An Tang – Université de Montréal

Co-investigators

• Julien Cohen-Adad – Polytechnique Montréal
• Martin Vallières – McGill University
• Robert Avram – Université de Montréal
• Julie Hussin – Université de Montréal
• Yohann Chiu – Université de Sherbrooke
• Angel Ruiz – Université Laval
• Sylvie Lambert – McGill University
• Esli Osmanlliu – McGill University

The widespread adoption of generative artificial intelligence (AI) tools is rapidly transforming the mental health landscape, creating both serious risks and significant opportunities. On the risk side, extended interactions with AI chatbots have been implicated in suicide and have led some users to develop what has been termed AI-related psychosis. With hundreds of millions of people using AI daily, even a small proportion of vulnerable users experiencing harm translates into suffering for thousands of individuals and their families. We will study who is at risk for AI-related mental health harms, and will develop practical safeguards to detect and interrupt harmful interaction patterns before they escalate. At the same time, AI presents opportunities for mental health care. Mental health remains among the most under-resourced areas of healthcare globally, and AI could dramatically expand capacity for both treatment and population-level screening. We will investigate how AI can be responsibly integrated into mental health treatment by identifying priorities, mapping the technical advances needed, and studying implementation in clinical settings. We will also examine how general-purpose commercial AI tools could be deployed to improve mental health screening and service access at a population level, particularly for underserved communities. This cluster brings together clinical psychiatrists, computational neuroscientists, AI researchers, ethicists, and lived-experience advocates to produce the first comprehensive, evidence-based account of AI’s impact on mental health — both the harms it can cause and the benefits it can deliver — and to develop the scientific infrastructure needed to maximize benefit and minimize harm.

Principal investigator

• David Benrimoh – McGill University

Co-Principal Investigators

• Lena Palaniyappan – McGill University
• Vera Bekes – McGill University
• Alban Elias Voppel – McGill University

Co-investigators

• Catherine Duclos – Université de Montréal
• Ian Jeffrey Gold – McGill University
• Vincent Taschereau-Dumouchel – Université de Montréal

The current AI landscape sees companies releasing increasingly capable AI models faster than we can even identify their risks, let alone mitigate them. Compounding matters, the latest models are agentic: they interact directly with web tools and user data (e.g., coding agents), and plan and execute complex tasks. AI agents raise the stakes on safety: as their capabilities grow, their failure modes become more subtle and harder to predict. Unsurprisingly in this context, Canadians show low rates of trust in AI. The key to safe and trustworthy AI systems lies in being able to draw reliable conclusions about the capabilities of agents (the science of evaluation), gaining insights into how agents solve tasks (the science of interpretability), and using this knowledge to mitigate agent’s unwanted behaviors (the science of robustness). Despite abundant research efforts in the AI safety community on interpretability and evaluation, existing methodologies have failed to produce findings that generalize to new contexts, and the cycle of finding a vulnerability and addressing it with specific guardrails has not led to broader robustness. While AI agents are new, these challenges are not. Our team has expertise spanning causality, reinforcement learning, and robotics, where we have developed theory around interpreting complex data, experimental design that supports reliable inferences, and making models robust to distribution shifts. Building on this foundation, the mission of this project is to bridge the gap between this theory and the complexity introduced by AI agents. We strive to make AI agents safer and more trustworthy by developing reliable methods that interpret, evaluate and improve the robustness of these systems.

Principal investigator

• Dhanya Sridhar – Université de Montréal

Co-Principal Investigators

• Simon Lacoste-Julien – Université de Montréal

Co-investigators

• Liam Paull – Université de Montréal
• Audrey Durand – Université Laval
• Golnoosh Farnadi – McGill University
• Gauthier Gidel – Université de Montréal

RAISE is a Québec research cluster bringing robust, reasoning and responsible artificial intelligence (AI) to the full agri-food value chain, from farm to fork. Québec’s food-processing sector is the province’s largest manufacturing employer, yet it is dominated by small and medium-sized enterprises (SMEs) whose data are scarce, heterogeneous, and proprietary, and whose decisions affect not only productivity and sustainability but also food safety, nutrition and public health. These conditions challenge mainstream AI and leave most actors unable to benefit from it. Anchored in INAF, the largest food and nutrition research institute in Canada, and its extensive network of industry and institutional partners, RAISE unites eight researchers across four universities spanning food production, food contaminants, nutrition, preventive health, animal science, co-product valorization, the microbiome, AI safety, and AI ethics. The cluster follows a three-tier strategy: (1) identify sector needs through a living, industry-co-designed observatory; (2) deliver deployable AI solutions for contaminant and pathogen detection, quality grading, shelf-life prediction, traceability, waste valorization, and nutritional quality optimization; and (3) advance long-term foundations — multimodal foundation models for food, digital twins of supply chains, precision-nutrition biomarkers, and a responsible-AI framework combining safety verification with ethical analysis. By transforming a fragmented, SME-dominated sector into a connected, data-driven ecosystem, RAISE aims to make food safer, more traceable, sustainable and health-promoting while generating generalizable advances in responsible AI for low-data, safety-critical domains. The cluster will train a new generation of researchers bridging AI, food, and health sciences while ensuring that the benefits of AI reach smaller and regional producers, as well as vulnerable populations.

Principal investigator

• Davoud Torkamaneh – Université Laval

Co-Principal Investigators

• Julien Parinet – Université Laval
• Daiva Nielsen – McGill University
• Elsa Rousseau – Université Laval

Co-investigators

• Pablo Valdés-Donoso – Université de Montréal
• Laurent Bazinet – Université Laval
• Frédéric Raymond – Université Laval
• Clayton Peterson – Université du Québec à Trois-Rivières

Supply chains and mobility systems underpin Canada’s economy and quality of life, yet they are strained by disruptions, inefficiencies, and environmental pressures. Decisions in these systems must be made under uncertainty about demand, travel and lead times, capacity, and human behaviour, using data that is noisy, incomplete, and often influenced by the decisions themselves. This Research Cluster develops a new generation of AI-enabled decision-support methods that integrate uncertainty quantification, machine learning, and optimization in an end-to-end fashion. Its scientific program is organized around three complementary axes: (1) uncertainty quantification designed for its downstream impact on decisions; (2) scalable methods for large-scale and multi-stage decision-making under uncertainty; and (3) responsible and human-centric AI decision support that is explainable, fair, privacy-aware, and robust to manipulation. Drawing on expertise in operations research, machine learning, statistics, transportation, and supply chain management across all five IVADO universities, the team validates its methods using real-world data through partnerships with leading transportation, logistics, energy, and mobility organizations. Anticipated outcomes include methodological advances published in leading scientific venues; open-source software and benchmarks; decision-support tools that improve the efficiency and resilience of Canadian supply chains and mobility systems, reduce their carbon intensity, and enhance services for groups facing systemic barriers; and a diverse cohort of highly qualified personnel equipped to accelerate the responsible adoption of AI in these nationally strategic sectors.

Principal investigator

• Erick Delage – HEC Montréal

Co-Principal Investigators

• Yossiri Adulyasak – HEC Montréal
• Jean-François Cordeau – HEC Montréal
• Emma Frejinger – Université de Montréal

Co-investigators

• Maryam Daryalal – HEC Montréal
• El Mehdi Er Raqabi – Université Laval
• Lijun Sun – McGill University
• Thibault Vidal – Polytechnique Montréal

Biodiversity is being lost faster than we can measure it. Our knowledge of where species live is full of gaps, biased toward a few well-studied places and organisms, and quickly outdated as the climate changes. Meanwhile, the tools we use to observe nature (satellites, drones, ground sensors, and citizen-science apps) are powerful but developed in isolation, and none can tell us where we should look next to learn the most. This Research Cluster builds a new kind of artificial intelligence for biodiversity: a “closed observation loop.” AI models will combine images and data across scales, from satellites to drones to in-the-field sensors and citizen-science photos, into robust representations of life that stay reliable in new places, for new species, and under a changing climate. Crucially, these models will quantify their own uncertainty and use it to decide where to look and what to sense next, steering drone flights, sensor placement, and citizen-science effort toward the observations that most improve our understanding, so this knowledge can guide decisions and actions to better manage, conserve, and restore ecosystems. We will develop and test these methods with partners who need them: monitoring vegetation, invasive plants, and species at risk with Hydro-Québec and Environment Canada; detecting invasive forest pests with federal agencies; tracking ecosystem recovery after mining; monitoring insects through automated AI sensor networks with the Montreal Insectarium; mobilizing citizen scientists to fill biodiversity knowledge gaps with iNaturalist; and, in full partnership with Emberá communities in Panama, mapping the giant trees that store immense carbon and biodiversity under Indigenous data governance. By making AI that is robust, reasoning, and responsible, the Cluster will help Québec and Canada meet their biodiversity and climate targets, empower communities and decision-makers, and train the next generation of researchers at the interface of AI and the environment.

Principal investigator

• Etienne Laliberté – Université de Montréal

Co-Principal Investigators

• David Rolnick – McGill University
• Laura Pollock – McGill University

Co-investigators

• Johanne Pelletier – Université de Montréal
• Christopher Pal – Polytechnique Montréal
• Tegan Maharaj – HEC Montréal
• Guillaume Blanchet – Université de Sherbrooke

Foundation models are AI systems trained on large and diverse datasets that can be adapted to many different questions. This project will develop and evaluate a new generation of foundation models for neuroscience, designed to connect measurements from different species across brain regions, neuroimaging modalities, electrical recordings, behaviour, and levels of organization ranging from molecules and cells to the whole brain. Building on the shared infrastructure, datasets, software, and expertise created by the current IVADO neuroAI Research Cluster, the project will pursue three goals. First, it will develop models that reveal common principles of brain organization across scales and species. Second, it will test whether these models can be adapted reliably to new research and clinical settings, including mental health, neurotechnology, and the study of cognition and brain disorders. Third, it will use neuroscience to improve AI itself. The team has expertise in evaluating how closely AI models align with humans, both in behaviour and in brain activity, and in identifying useful inductive biases, which are built-in learning principles that guide how models are trained. A dedicated technical team will provide secure infrastructure, standardized workflows, pretrained models, evaluation tools, training, and hands-on support. This will make advanced AI methods more accessible to researchers who do not have specialized expertise or major computing resources, while promoting reproducibility, responsible data use, and equitable participation. The project will bring together researchers in AI, fundamental and clinical neuroscience across Québec, Canada, and international networks. Its long-term legacy will include reusable tools and models, a highly trained interdisciplinary workforce, stronger links between fundamental research and clinical applications, and new brain-inspired approaches to building more reliable and adaptable AI.

Principal investigator

• Flavie Lavoie-Cardinal – Université Laval

Co-Principal Investigators

• Karim Jerbi – Université de Montréal
• Lune Bellec – Université de Montréal

Co-investigators

• Pouya Bashivan – McGill University
• Eva Portelance – HEC Montréal
• Guillaume Dumas – Université de Montréal
• Catherine Duclos – Université de Montréal
• Shahab Bakhtiari – Université de Montréal
• Guillaume Lajoie – Université de Montréal

This research cluster aims to develop AI methods to enable energy-e`icient, ubiquitous, and ultra-reliable communication networks, termed “sustainable mission-critical (SMC) networks,” that can support critical systems. It also considers in depth the application of such an SMC network to active power distribution systems, which are an essential component of next-generation power grids required to support the largescale integration of renewable power generation. The central element of the proposal is the development of trustworthy and energy-e`icient AI models specifically designed to operate the SMC network. The project is divided into three themes: AI methods for SMC networks, AI-based signal processing, and AIdriven communication for active power distribution networks. The ambition of the proposed research cluster is to establish an international reference on AI-native SMC networks and its application to sustainable power distribution, achieved through collaborations between experts with complementary expertise in communication systems, machine learning, information theory, and dynamical systems.

Principal investigator

• François Leduc-Primeau – Polytechnique Montréal

Co-Principal Investigators

• Gunes Karabulut-Kurt – Polytechnique Montréal
• Antoine Lesage-Landry – Polytechnique Montréal
• Jennifer Tang – Polytechnique Montréal

Co-investigators

• Georges Kaddoum – École de technologie supérieure
• Bassant Selim – École de technologie supérieure

AI is transitioning from isolated models to ecosystems of interacting agents at a pace that far exceeds our scientific understanding of how to design, analyse, and govern them. Agentic systems are entering clinical workflows and emergency response, while the field still lacks understanding of when a multi-agent architecture helps. Our research cluster proposes to build the science of how to measure the robustness of multi-agent systems in the foundation model era, focusing on one core concept that subsumes the challenges of multi-agent systems: asymmetry. This asymmetry can be in information, capability, compute, and incentives, and it is the key differentiator between a single (very large) system and a truly multi-agent system. This scientific foundation will enable building more robust and reliable agentic AI deployments.

Principal investigator

• Gauthier Gidel – Université de Montréal

Co-Principal Investigators

• Aaron Courville – Université de Montréal
• Doina Precup – McGill University

Co-investigators

• Aditya Mahajan – McGill University
• Audrey Durand – Université Laval
• Christopher Pal – Polytechnique Montréal
• Dhanya Sridhar – Université de Montréal
• Laurent Charlin – HEC Montréal
• Siva Reddy – McGill University

From electoral fraud claims to climate denial to vaccine scepticism, many of Canada’s hardest challenges today are information problems, and generative AI threatens to make them intractable by producing highly convincing content at scale, whether true or false. Understanding how threats to information integrity affect AI systems and society requires interdisciplinary expertise in machine learning, automated verification, media ecosystems, human cognition, public policy, and scientific integrity. These six themes form the core of the proposed Information Integrity Research Cluster, anchored by the deployed fact-checking application Veracity, mentioned in the 2026 Canadian AI Strategy, and by our 2027 IVADO thematic semester on this topic.
The cluster has six axes. Detection & Verification advances automated fact-checking and deepfake detection through Veracity, with systems that recognize reliable evidence. Data & Model Foundations studies the provenance of training data and develops methods to make generative models more factual and resistant to manipulation. Platform & Ecosystem Dynamics monitors how misinformation and coordinated manipulation spread. Human Cognition, Persuasion & Debunking studies why people believe and share misleading content, and how false beliefs can be corrected. Governance, Policy & Law examines the regulatory frameworks needed to sustain a trustworthy information environment. Scientific Integrity assesses the production and reliability of the scientific evidence that verification depends on.
Information integrity is a necessary condition for robust, reasoning, and responsible AI. A model trained on unreliable data is neither robust nor trustworthy. Likewise, the reasoning abilities of AI are limited if its internal knowledge base is polluted by low-quality information. Finally, the responsible alignment of AI with human values depends on the diversity of a model’s training data and the credibility of the external sources it retrieves.

Les agents d’intelligence artificielle représentent une nouvelle génération de systèmes capables d’agir de manière autonome, de collaborer avec les humains et de réaliser des tâches complexes. Leur
déploiement transformera le travail, les organisations, les services publics et les interactions avec les citoyens. Cette autonomie croissante soulève toutefois une question centrale : dans quelles conditions les organisations peuvent-elles déployer des agents d’IA de manière éthique, responsable et digne de confiance? Ce projet réunira une équipe interdisciplinaire afin d’étudier cette question dans des contextes réels de déploiement. Il examinera trois dimensions complémentaires. Sur le plan technique, il étudiera les caractéristiques permettant de développer des agents d’IA robustes, sécuritaires, explicables et auditables, ainsi que la traduction d’exigences sociales, culturelles, juridiques et éthiques en exigences techniques vérifiables. Sur le plan organisationnel, il analysera les transformations du travail et les adaptations nécessaires des mécanismes de gouvernance, d’éthique, de supervision et de responsabilité face à des systèmes plus autonomes. Sur le plan sociétal, il étudiera les conditions d’adoption et de légitimité des agents d’IA, les résistances à leur déploiement et leurs eoets sur dioérentes populations, notamment les populations vulnérables et les Premiers Peuples. Réalisé en collaboration avec des organisations publiques, privées, communautaires et autochtones, le projet vise à développer un Cadre canadien d’évaluation du déploiement responsable des agents d’IA. Celui-ci permettra de mieux guider les organisations et les décideurs quant aux conditions de déploiement, aux niveaux appropriés d’autonomie et de supervision humaine, ainsi qu’aux mécanismes nécessaires pour favoriser une utilisation éthique, responsable et socialement légitime des agents d’IA.

AI can now generate hypotheses, code, and papers, making verification the new bottleneck of science. We reframe verification as a reasoning problem: decomposing claims, grounding them in the literature and in a paper’s own code and data, and knowing what cannot be verified — fundamental capabilities current AI systems (large language models or LLMs in particular) lack. We couple these verifiers with a scientific discovery engine whose outputs must pass verification, forming a human-operated loop, and run human studies of how scientists trust and oversee it. Our north star: by 2030, an open verifier whose reviews of computer science and mathematics papers match expert scrutiny, deployed in a live venue.

Principal investigator

• Laurent Charlin – HEC Montréal

Co-Principal Investigators

• Hugo Larochelle – Université de Montréal
• Christopher Pal – Polytechnique Montréal
• Siva Reddy – McGill University

Co-investigators

• Aishwarya Agrawal – Université de Montréal
• Ian Arawjo – Université de Montréal
• Oana Balmau – McGill University

Physical AI is poised to have a massive impact on our lives and business in the near future. This optimism stems from the rapid advances in the capabilities of more general-purpose robots in recent months. These advances have been driven by narrow automation strategies. Development of large-scale (by robotics standards) datasets has served as a backbone for imitation learning strategies such as vision-language action (VLA) models. Increasingly high-fidelity and data-driven simulation environments have enabled policy learning through reinforcement learning. Nevertheless, the number of real-world deployments of this technology that are actually providing commercial value is still extremely limited. Instead, demonstrations are done in sanitized lab environments that do not well reflect the real-world operating conditions in which they would be deployed. Furthermore, the requirements for robustness and safety are severe if these technologies are meant to work alongside people. In this proposed research cluster, we will attempt to overcome these challenges so that physical AI systems can be deployed in real manufacturing environments and provide real economic and social value. The key insight that we will exploit is that most manufacturing settings are, at least somewhat, controlled. This renders the unconstrained “Generalist AI” approach ill-suited since it lacks the capacity to exploit the reduced variability in the operating environment in which it operates. We will develop methods that enable us to flexibly impose such constraints to allow single and multiple physical AI agents to work alongside people in small and medium-sized factories in a way that is safe, robust, responsible, and efficient. The specific sector that we will target is manufacturing for low-cost, environmentally friendly housing – a sector of high emphasis for the Canadian government.

Principal investigator

• Liam Paull – Université de Montréal

Co-Principal Investigators

• Christopher Pal – Polytechnique Montréal
• AJung Moon – McGill University
• Glen Berseth – Université de Montréal

Co-investigators

• James Forbes – McGill University
• Giovanni Beltrame – Polytechnique Montréal
• Pierre-Yves Lajoie – Polytechnique Montréal
• Jana Pavlasek – Polytechnique Montréal
• David Meger – McGill University
• Claire Dune – Université Laval

Artificial intelligence (AI) and quantum information science are two of the most transformative areas of modern research, yet they have largely advanced in parallel. This cluster brings them together around a simple idea: each field can accelerate the other. Quantum computers can represent and process information in ways classical machines cannot, potentially enriching what AI systems can learn and infer; in turn, AI can make quantum hardware more usable, reliable, and scalable. Uniting researchers in AI, quantum algorithms, physics, engineering, and medicine across Quebec’s five leading universities (the U5) and its wider quantum ecosystem, the cluster pursues this “virtuous circle” with rigour rather than hype, guided by IVADO’s R3AI mission of robust, reasoning, and responsible AI. A central goal is to determine, and to prove, exactly when quantum resources give a genuine advantage for real AI and scientific-computing tasks, and when they do not. Rather than overpromising, the team builds the strongest possible classical methods against which every quantum claim is tested, so that only certified advantages survive. Anticipated benefits include faster and more trustworthy tools for molecular science and drug discovery, lower energy costs for AI, open benchmarks the community can adopt, and a new generation of scientists fluent in both quantum information and machine learning. The cluster also treats equity and inclusion as a design goal, lowering the high barriers to entry in quantum computing through open tools, shared training, and broad recruitment, so that the first quantum-enabled AI cohort reflects the diversity of the communities it serves.

La Stratégie « IA pour tous » du gouvernement fédéral s’appuie sur des mesures législatives (pilier 1) et des subventions (piliers 1 à 5), sans plus d’explications. Or, les choix de politiques publiques (macro) doivent reposer sur des microfondements solides, soit une compréhension fine des implications et conditions de l’adoption de l’IA par les individus (micro) et des répercussions pour les organisations (meso). Notre regroupement interdisciplinaire étudie les microfondements de l’IA responsable à travers les liens entre l’agentivité individuelle (micro), l’encadrement organisationnel (meso) et l’action étatique (macro). Suivant une démarche ascendante, des enjeux vécus par les individus (littératie IA, rapports aux systèmes algorithmiques), en passant par la gouvernance des organisations (redevabilité, frugalité, relations avec les fournisseurs de modèles de fondation), jusqu’aux instruments publics (régulation, marchés publics, souveraineté numérique), le programme combine recherches expérimentales en laboratoire, enquêtes et études économétriques, études de cas organisationnelles, analyses juridiques et sociales, et ateliers avec les citoyens et les décideurs publics. Nos recherches examinent les composantes de ces niveaux et les interactions entre eux, chacun servant de fondement aux niveaux supérieurs – et réciproquement, de contrainte aux niveaux inférieurs. Cette démarche permet d’élucider l’architecture reliant ces niveaux, et les moyens d’atteindre et de préserver une IA responsable. Fort des acquis du Regroupement 4 de la première phase d’IAR3, nous en prolongeons les travaux. Nous produisons à la fois de nouvelles connaissances scientifiques et des outils utilisables : pour les individus, des guides de littératie IA et de contestabilité ainsi que des principes de conception préservant le jugement humain; pour les organisations, des canevas d’audit et des clauses contractuelles; pour l’État, des bacs à sable réglementaires, des clauses de marchés publics, et des lignes directrices et meilleures pratiques.

Principal investigator

• Pierre Larouche – Université de Montréal

Co-Principal Investigators

• Hassane Alami – Université de Montréal
• Foutse Khomh – Polytechnique Montréal
• Lyse Langlois – Université Laval

Co-investigators

• Catherine Beaudry – Polytechnique Montréal
• Kyle Conway – University of Ottawa
• Vincent Denault – Université de Montréal
• Julie Alice Gramaccia – University of Ottawa
• Marion Korosec-Serfaty – Université du Québec à Montréal
• Thang Le Dinh – Université du Québec à Trois-Rivières
• Pierre-Majorique Léger – HEC Montréal
• Véronique Nabelsi – Université du Québec en Outaouais
• Houari Sahraoui – Université de Montréal

Modern AI and Large Language Models (LLM) technologies are trained on massive text corpora composed primarily of American English and prestige dialect content. Consequently, they perform poorly for minority languages, regional dialects, and culturally distinct communities. This can limit access to AI tools and contribute to the erosion of linguistic and cultural diversity. Our research cluster aims to develop language technologies that better serve low-resource languages and dialects while respecting the cultures of the communities that use them. The project will focus on three priorities: Canadian regional dialects, Indigenous languages of the Americas, and underrepresented international languages. Working in partnership with communities, researchers will develop ethical approaches to data collection, create resource-efficient models, design new methods for evaluating cultural and linguistic accuracy, and build practical tools to benefit the communities. A central principle of the project is that communities should have a meaningful role in decisions about their language data, cultural knowledge, and digital resources. The initiative is grounded in the principles of responsible AI, including inclusion, cultural preservation, transparency, and respect for community ownership of data. By helping AI understand a broader range of languages and cultural contexts, the project seeks to reduce digital inequalities and ensure that the benefits of AI are accessible to more people. Led by researchers from Université Laval, Université de Montréal, McGill University, and UQAM, the project brings together expertise in artificial intelligence, linguistics, speech technology, and community-based research. Its outcomes will support language learning and preservation, strengthen cultural representation in AI, and create opportunities for communities, organizations, and businesses to participate more fully in the digital economy.

Principal investigator

• Richard Khoury – Université Laval

Co-Principal Investigators

• Ayla Rigouts Terryn – Université de Montréal
• David Adelani – McGill University

Co-investigators

• Fatiha Sadat – Université du Québec à Montréal

Despite extraordinary advancements in Artificial Intelligence (AI), current systems remain fundamentally static: trained offline on fixed datasets and deployed without the capacity to learn from ongoing experience. As real-world environments, human preferences, and language evolve, static models rapidly become outdated. Frequently retraining massive foundation models from scratch is computationally, financially, and environmentally unsustainable.
This research program establishes Continual Learning (CL) as the core paradigm for next-generation AI. By transitioning machine learning from static deployment to lifelong adaptation, this initiative will design systems that continually acquire knowledge, refine skills, and adapt to non-stationary environments, much like humans do. The scientific program is structured around four integrated research axes:
• Optimization & Algorithms: Developing optimization methods that prevent catastrophic forgetting while preserving model plasticity during continual pre-training.
• Architectures & Dynamic Memory: Designing dynamic modular structures and memory consolidation mechanisms to transfer context into long-term knowledge.
• Continual Decision-Making: Advancing continual reinforcement learning and world models to enable autonomous agents to learn safely in dynamic environments.
• Continual Alignment & Safety: Developing lifelong preference optimization to ensure deployed models remain trustworthy, safe, and aligned over time.
Bringing together leading experts across Canadian and international institutions, this regroupement will solidify Canada’s global leadership in AI. By replacing costly retraining with continuous adaptation, this work will dramatically reduce AI’s carbon footprint, democratize model adaptation for startups and academia, and deliver reliable, adaptive AI across healthcare, robotics, and natural language processing.

Confrontés aux tensions géopolitiques, à la reconfiguration des chaînes de valeur, à la transition énergétique et à la transformation numérique, les gouvernements et les entreprises doivent décider dans des environnements toujours plus incertains. Ce regroupement développera une plateforme d’intelligence artificielle qui transforme des données géospatiales, économiques, logistiques et textuelles en prévisions, simulations et scénarios d’aide à la décision. S’appuyant sur le jumeau numérique économique de la région du Saint-Laurent–Grands Lacs (GVCdtLab), soutenu par le Fonds national des corridors commerciaux (FNCC) de Transports Canada, il développera une nouvelle génération de modèles d’IA robustes, raisonnants et responsables, représentant à la fois la structure physique de l’économie — territoires, infrastructures, chaînes de valeur — et le comportement de ses agents — attentes, décisions, politiques. Le premier volet mobilisera l’IA géospatiale, les graphes et l’apprentissage multimodal pour cartographier les réseaux économiques, détecter les dépendances critiques, simuler les effets de perturbations économiques, climatiques ou géopolitiques et éclairer les stratégies de diversification commerciale et d’investissement. Le second volet combinera apprentissage automatique, économétrie, analyse automatisée de textes et IA générative pour produire des indicateurs avancés de l’activité économique, améliorer les prévisions macroéconomiques, mesurer et simuler les attentes des ménages et des entreprises, et évaluer ex ante les effets des politiques économiques et budgétaires. Les outils développés appuieront des décisions publiques et privées plus éclairées et plus équitables — incluant l’analyse des effets différenciés des chocs selon les groupes et les territoires —, contribueront à former une relève hautement qualifiée et feront du Québec un pôle de calibre international en IA pour la décision économique.

Principal investigator

• Valérie Bélanger – HEC Montréal

Co-Principal Investigators

• Kevin Moran – Université Laval
• Martin Trépanier – Polytechnique Montréal

Co-investigators

• Thierry Warin – HEC Montréal
• Dalibor Stevanovic – Université du Québec à Montréal
• Aristide Houndetoungan – Université Laval
• Marcelin Joanis – Université de Montréal
• Nathalie de Marcellis-Warin – Polytechnique Montréal

Antibiotics are among the most important tools in modern medicine, but many bacteria are becoming resistant to existing drugs, underscoring the urgent need for new antibiotics. Finding them is slow because scientists must search through enormous numbers of possible molecules, synthesize the most promising ones, test whether they affect bacteria, and understand how they work. This project will develop new artificial intelligence methods to make antibiotic discovery smarter and faster. Instead of asking AI only to predict which molecule might work, we will build a “make–test–reason” system. The AI will suggest molecules that can actually be synthesized, select experiments most likely to teach us something useful, learn from the results, and then improve its next suggestions. Antibiotic discovery is an ideal test case for methods that can be broadly applied to drug discovery because bacteria can be tested rapidly in the laboratory. The project will combine expertise in AI, chemistry, microbiology, and drug discovery. Promising molecules will be synthesized using rapid chemistry and tested against bacteria. The results, including successes, failures, and uncertain findings, will be fed back into the AI system to improve its accuracy. The goal is not only to find better starting points for new antibiotics, but also to create AI tools that explain why a molecule may work, what bacterial processes it may affect, and whether resistance could emerge. The project will also train researchers who can work across AI, chemistry, and biology, and will develop reusable methods and benchmarks that will help other areas of drug discovery.

Principal investigator

• Yves Brun – Université de Montréal

Co-Principal Investigators

• Audrey Durand – Université Laval
• Alex Hernandez-Garcia – Université de Montréal
• Olivier Mailhot – Université de Montréal

Co-investigators

• Anne Marinier – Université de Montréal
• Sarath Chandar – Polytechnique Montréal
• Sébastien Lemieux – Université de Montréal
• Chris Maddison – University of Toronto