Impact Report 2024-2025

Refgen, a generative pipeline combining machine learning with physics-based principles, was developed by Mila researchers Pierre-Luc Bacon, Adrien Goldszal, Diego Calanzon and Vincent Taboga to accelerate the discovery of safer, more efficient alternatives to current refrigerants, most of which are potent greenhouse gases. With this AI-powered system, molecular generation is fine-tuned to ensure chemical validity while incorporating predictive models of critical properties, equations of state, thermochemical behaviour, and full vapour-compression cycle performance. By enabling reinforcement learning under thermodynamic constraints, the Refgen team hopes to discover molecules that are both reliable and environmentally sustainable. Embedding physics directly into the learning process allows the system to leverage limited data and explore beyond the few hundred known refrigerants, aiming to uncover entirely new compounds with enhanced performance and reduced environmental impact. 

PRISM, an innovative generative AI model developed by Mila researchers Tal Arbel, Amar Kumar and Anita Kriz, was created to enhance the interpretability and trustworthiness of medical imaging. Unlike traditional "black-box" AI systems, PRISM utilizes Stable Diffusion, a vision-language foundation model, to generate high-resolution counterfactual images. These images depict alternative scenarios, such as the absence of a specific disease, by precisely editing medical images while preserving other attributes. This capability allows clinicians to understand the rationale behind AI-generated diagnoses, fostering greater confidence in AI-assisted decision-making. PRISM's design emphasizes accessibility, enabling interaction through natural language prompts, and adaptability, making it suitable for various medical imaging applications. By providing clear visual explanations, PRISM aims to bridge the gap between AI predictions and clinical reasoning, paving the way for more transparent and reliable AI tools in healthcare

This groundbreaking study, conducted by Mila researchers Praneet Suresh, Jack Stanley, Sonia Joseph, Luca Scimeca and Danilo Bzdok, introduced a novel method to address AI hallucinations. Traditional safety measures focus on filtering inputs and outputs, but this approach often overlooks the internal processes leading to hallucinations. The team discovered that when fed with degraded or meaningless inputs, AI models activate internal "concept units" to compensate, effectively "inventing" structure and meaning. By identifying and neutralizing these internal origins of hallucinations, false outputs can be prevented before they reach the end user. This advancement empowers organizations designing LLM-powered products and services to build more reliable, trustworthy AI systems, accelerating the safe and widespread adoption of AI across all industries.

Created by Mila researchers Gabriel Tseng and David Rolnick, Galileo represents a groundbreaking shift in Earth observation, making satellite and human activity data accessible to anyone. This AI model uses self-supervised learning to teach itself by discovering patterns across vast archives of unlabelled data from decades of Earth monitoring. From climate tracking to land cover mapping, Galileo enables researchers, policymakers, and communities to gather and interpret publicly accessible data—to better understand our planet's health metrics and to act on alarming signals. By seamlessly processing both spatial and temporal data, Galileo reveals how landscapes evolve over time. Applications span wildfire risk monitoring, global agriculture mapping, and on-the-ground response—such as quickly generating high-quality cropland maps for entire countries, a task that previously took months.

In April 2025, Mila, in collaboration with Université Laval and Calcul Québec, launched TamIA, Canada's inaugural AI computing cluster tailored for academic research. Located on Université Laval's campus, TamIA is the first operational component of the Pan-Canadian AI Compute Environment (PAICE). This state-of-the-art infrastructure is designed to support a wide range of AI research, including applications in healthcare, environmental science, biodiversity, robotics, and language processing. Once fully operational, TamIA will comprise 75 interconnected servers, 4,000 processor cores, and 38,000 gigabytes of RAM, providing exceptional computational power for researchers across Quebec and Canada. 

Mila developed Milabench, a transparent graphics processing unit (GPU) evaluation tool that helps data centres select the most suitable hardware for AI research tasks. Unlike commercial solutions, Milabench provides exhaustive and unbiased performance-cost analysis, informed by nearly 1,000 Mila-affiliated scientific articles published in 2023. The tool was showcased at SuperComputing and at the The Future of Benchmarks in Supercomputing workshop hosted at ISC High Performance, highlighting its role in advancing standardized benchmarking practices. Expanding beyond Mila, Milabench has been adopted by The Vector Institute, the Digital Research Alliance of Canada, and the MINERVA consortium (regrouping several major European compute centres), enabling organizations worldwide to make informed hardware choices and optimize AI research outcomes. 

In 2025, Mila’s Entrepreneurship Lab earned official recognition from Quebec’s Ministry of Economy, Innovation and Energy (MEIE) as an accredited incubator and accelerator for high-growth, innovative startups. Launched in 2022, the ELab has supported the growth of over 50 startups, which have collectively raised more than $100 million. In the past year alone, the ELab enabled the creation of 12 Mila-affiliated startups, working across a variety of fields. In collaboration with Montréal International, the ELab also launched the Venture Up North Residency, a unique initiative to welcome AI-focused startups from U.S. universities to Montréal in the summer of 2025, while also laying the groundwork for a permanent residency to attract global talent.

In October, Mila’s AI Ventures program, in collaboration with Cohere, Radical Ventures, and Amiral, hosted Rise of the Venture Scientist, an event that brought together Canada’s leading AI researchers, venture scientists, and investors. Attendees gained exclusive insights into the future of venture creation at the intersection of science and entrepreneurship. Key speakers included Joelle Pineau, Chief of AI at Cohere, and Mila’s newly appointed Scientific Director Hugo Larochelle, alongside other visionary leaders driving AI innovation in Canada. 

With financial support from the National Research Council of Canada’s Industrial Research Assistance Program (NRC IRAP), Mila provided consulting services to more than 40 innovative Canadian SMEs working on applied machine learning projects. These collaborations enable companies to successfully integrate machine learning into their operations, strengthen in-house capabilities, and advance their innovation journey. For instance, Halifax-based Densitas and Mila collaborated on an AI solution to improve breast cancer screening and advance women's health technology. Meanwhile, Metareal, based in Magog, worked with Mila to develop a Vision Transformer (ViT)-based model that automatically corrects image alignment in 3D virtual tours, enabling the company to launch its first AI-driven product.

Mila joined forces with 5C and Hypertec in a strategic collaboration to launch a $250 million Sovereign AI Research Hub and a new campus in LaSalle. Mila will leverage its expertise to provide researchers and startups with secure access to advanced computing resources, accelerating AI adoption and promoting sustainable infrastructure design. This landmark partnership was announced at ALL IN, Canada’s largest AI event, with key government officials, including Mélanie Joly, Minister of Industry and Minister responsible for Canada Economic Development for Quebec Regions, Evan Solomon, Minister of Artificial Intelligence and Digital Innovation, and Christine Fréchette, Quebec Minister of Economy, Innovation and Energy. 

In April, Mila hosted its first AI & Climate Talks, bringing together researchers, industry leaders, government agencies, and non-profits to explore how AI can drive climate action. Discussions focused on applications in energy systems, biodiversity conservation, and climate finance, while emphasizing the importance of research-industry collaboration and the power of open science. Through panels, presentations, and informal discussions, the event highlighted how machine learning is already being applied across sectors to tackle urgent climate challenges.

Mila startup Rubisco AI played a key role in developing the canopy tree mapping technology that helped the Limelight Rainforest team win the $10M XPRIZE Rainforest competition. The team earned the prize for their remarkable “ability to survey 100 hectares of tropical rainforest in 24 hours and deliver the most impactful real-time insights within 48 hours.” Founded by Étienne Laliberté (Associate Academic Member of Mila) and postdoc Arthur Ouaknine, Rubisco AI leverages advanced AI and computer vision to convert drone imagery into precise, tree-level data on carbon storage and biodiversity. 

In 2025, Mila launched three new studios—dedicated to AI for Humanity, AI for Climate, and AI Safety—marking a strategic effort to connect cutting-edge AI research with tangible community impact. Grounded in applied expertise in engineering and design, these studios focus on delivering impactful, end-to-end AI projects that accelerate responsible adoption and address critical societal needs, while building a reusable body of AI artifacts and open-source tools that amplify our collective impact. One such example is the AI Safety team’s first initiative, which aims to mitigate risks associated with the misuse of chatbots by users experiencing psychological distress. To complement the work of all three studios, Mila is also offering specialized consulting in responsible AI, positioning itself as a key partner for organizations seeking to implement AI ethically and securely. 

The Responsible AI (RAI) Advisory is strategically positioning Mila as a key partner for organizations seeking to implement AI both ethically and securely, accelerating the adoption of AI that is effective and responsible. This year, the RAI Advisory conducted comprehensive impact assessments to proactively identify and map potential unintended harms across a diverse range of AI-enabled products, delivering customized technical and procedural mitigation strategies. Key initiatives included supporting the City of Vaudreuil-Dorion in the trustworthy deployment of conversational agents for citizen inquiries, and collaborating with BrainBox AI to ensure safety and ethical alignment for virtual engineering assistants in the building automation industry. 

Released in January 2025, the International AI Safety Report 2025 serves as a global handbook for policymakers. Chaired by Yoshua Bengio and authored by 100 international experts, it establishes the first comprehensive scientific understanding of advanced AI systems and their risks. This first edition of the report highlights critical developments: the most advanced AI systems can write sophisticated code, identify cyber vulnerabilities, and operate as autonomous agents. These capabilities underscore three categories of AI risks: malicious use, system malfunctions, and systemic risks (such as workforce disruption, privacy concerns and environmental impacts). Crucially, the report emphasizes that while capabilities can advance at varying speeds, outcomes depend entirely on the choices societies and governments make today and in the future.

Mila launched its flagship AI Policy Fellowship in August 2025 to bridge the gap between AI research and other disciplines, and to inform sound AI policymaking. The program enables Fellows to collaborate with Mila researchers, developing insights that address societal challenges and opportunities in AI development, deployment, and governance. Over six months, each Fellow pursues a project tackling pressing questions—courts, addressing accountability in multi-agent interactions, and designing safer chatbots for democratic resilience and youth mental health. The fellowship has already produced its first policy note examining the intersection of quantum technology and AI, and will be delivering over 20 policy oriented events with key stakeholders of local and international AI ecosystems together with the inaugural cohort. Building on the previous year’s success, the AI Insights for Policymakers Program returned in 2025. This joint initiative of CIFAR and Mila provided evidence-based guidance to over 20 government and civil society organizations navigating complex AI policy challenges, connecting decision-makers with leading AI researchers. 

Canada was the country of honour at Vivatech in France, with Mila at the heart of the Canadian delegation. In June 2025, our 12-member team of experts and professionals engaged with global leaders including French President Emmanuel Macron and France’s first AI Minister, Clara Chappaz, and international organizations including UNESCO. The delegation showcased Canadian AI values and expertise through high-level roundtables with Canadian Minister Evan Solomon, meetings with Quebec Premier François Legault, bilateral discussions, and public debates—including a New York Times discussion on whether “AI is the new plastic.” Mila’s presence extended to diplomatic events at the Canadian Embassy and l’Élysée, positioning Canada as a thought leader in AI development.

Mila was delighted to welcome the Right Honourable Mark Carney, Prime Minister of Canada, and Evan Solomon, Minister of Artificial Intelligence and Digital Innovation for dynamic discussions on AI's potential to drive innovation, social progress, and economic resilience. The visit, held in August 2025, brought together key players from our vibrant AI ecosystem for substantive exchanges on the technology's transformative possibilities. The Prime Minister also witnessed firsthand how breakthroughs in AI and robotics are reshaping society, from cutting-edge research labs to direct conversations with students.

Mila’s Founder Yoshua Bengio met with Pope Leo XIV in Rome and participated in the World Meeting on Human Fraternity—a Vatican-sponsored summit featuring 15 thematic roundtables on critical issues shaping humanity's future. Held in September 2025, the event brought together religious, political, and scientific leaders to examine AI's ethical dimensions, emphasizing that technological progress must be guided by shared human values. At the Artificial Intelligence roundtable, Professor Bengio joined global experts to address the profound challenges posed by Artificial General Intelligence (AGI)—AI systems potentially capable of performing as well as most humans on most cognitive tasks. Discussions centred on major risks and the urgent need for robust regulation to ensure AI development serves humanity's collective good.

Mila reaffirmed its leadership in AI through seven presentations at ALL IN 2025, the most important AI event in Canada. Held in Montréal and convening more than 6,500 leaders and innovators, the gathering provided a platform for championing robust, responsible AI development aligned with societal needs. CEO Valérie Pisano discussed how AI drives social transformation with journalist Patrice Roy, while new Scientific Director Hugo Larochelle and Joelle Pineau, Associate Professor at McGill University and Core Academic Member at Mila, explored strategies for accelerating the translation of Canadian AI research into commercial applications—effectively bridging the gap between academic innovation and real-world impact. Yoshua Bengio addressed the critical importance of implementing safeguards for responsible AI development. Additional contributions examined AI's transformative potential across diverse sectors: Anna Jahn on the future of public administration, Shingai Manjengwa on the evolution of work, and Alex Hernández-García on leveraging AI to tackle major global challenges.

Mila’s AI Governance Team, in collaboration with IVADO, produced a guide on the responsible integration and governance of AI for Quebec’s Ministry of Higher Education. This guide is designed to empower cégeps and universities to better understand both the benefits and risks of AI adoption. It delivers concrete courses of action, including reflection frameworks and a self-assessment tool to measure institutional maturity. Key recommendations centre on: fostering responsible experimentation, ensuring community participation in governance, and adapting existing resources to maximize efficiency.

Our AI4Good Lab equips women and gender-diverse individuals with skills and networks to launch successful AI careers. Approaching its 10th anniversary, the program now boasts over 400 alumni across Canada. In 2025, AI4Good Lab engaged 74 participants, inspiring 91% of them to pursue AI careers while supporting 14 socially impactful machine learning projects. This year emphasized region-specific programming, with expertise tailored to local issues and problem-solving approaches. At the summer 2025 Demo Day, trainees presented their capstone projects addressing pressing challenges: urban accessibility, inclusive physical therapy, neglected tropical diseases, and climate change. These projects demonstrate how technical skills, combined with social consciousness, can drive meaningful innovation that benefits communities nationwide.

Mila convened the first-ever Indigenous AI Gathering, bringing together over 170 participants to explore the intersections of AI and Indigenous knowledge systems. Indigenous leaders, students, entrepreneurs, government representatives, and industry partners from across Canada and beyond engaged in rich interdisciplinary dialogue throughout the two-day event held in July 2025. The gathering showcased diverse perspectives on AI's future through multiple lenses. Participants in Mila's Indigenous Pathfinders in AI program presented their projects, while panel discussions examined critical themes such as AI ethics and Indigenous knowledges, inclusion in AI amid changing economic and political landscapes, AI and environmental stewardship, and Indigenous languages and AI. The collaborative event was organized by Mila, Abundant Intelligences, Indspire, and IVADO.

Through its new strategic learning partnership with the Information and Communications Technology Council (ICTC), Mila is helping accelerate responsible AI adoption in Canada. This collaboration unites ICTC’s expertise in workforce readiness, business innovation, and digital adoption with Mila’s global leadership in AI research and responsible innovation. Together, the organizations will help Canadian students, workers, and enterprises build the technical, ethical, and governance skills required to thrive in an AI-driven economy. The partnership will also support SMEs in adopting AI responsibly, while reaching underserved and equity-deserving communities via inclusive and bilingual training programs. Beyond capacity building and business support, the partnership will inform policy dialogue through joint research and ecosystem engagement on AI competitiveness, governance, and inclusive innovation.

Mila's annual Techaide charity event brought together hundreds of members from Montréal's tech community. Held in April 2025, the fundraiser generated over $110,000 to support Centraide of Greater Montréal's mission to fight poverty and social exclusion. The evening featured inspiring talks from leading AI researchers: David Ifeoluwa Adelani (Mila/McGill), Kyunghyun Cho (NYU), Golnoosh Farnadi (Mila/McGill), Sara Hooker (Cohere for AI), Tegan Maharaj (Mila/HEC Montréal), and Derek Nowrouzezahrai (Mila/McGill). Speakers explored critical topics spanning causal inference, equity and privacy in foundation models, and responsible AI in an era of increasingly complex systems.