The Flowers team or in praise of curiosity

How do we learn? This is the question at the heart of the work carried out by the Flowers AI & CogSci project team, led by Inria and the University of Bordeaux. “We are interested in the ability to learn new knowledge and skills throughout life,” explains Pierre-Yves Oudeyer, the team's director. “This is particularly true in complex environments such as those experienced by children. How do they manage to learn so much in such a limited time and with so few resources? We believe that curiosity is one of the key ingredients for achieving this.” 

To explore the different angles of this vast field of research, the team brings together a wide range of disciplines. Alongside its three permanent researchers—Pierre-Yves Oudeyer, Inria research director and specialist in artificial intelligence (AI) and cognitive science, Hélène Sauzéon, professor of psychology, and Cécile Mazon, lecturer in cognitive science—there are around twenty doctoral and postdoctoral students, as well as external partners from a wide range of fields: computer science, mathematics, psychology, neuroscience, educational technology, and more. 

Flowers' first area of research: developing models to better understand curiosity in humans. To do this, the team developed the “learning progress hypothesis.” The idea? “In our brains, neural circuits reward situations where we make progress, directing our learning towards those that are most rewarding in terms of satisfaction,” explains Pierre-Yves Oudeyer. This theory was conceived between 2003 and 2004, then developed theoretically and in simulation for almost two decades, and finally confirmed in 2021 by Flowers—and other teams around the world thereafter—thanks to new experimental devices that allow the study of curious exploration in humans.

Today, the team is testing the models obtained, in particular through cognitive psychology experiments combined with brain imaging, with its partners in the European DevCur program, which aims to track the development of curiosity and metacognition during adolescence. 

Another major advance linked to this work is the refinement and development of the concept of “autotelic curiosity” (from the Greek “auto,” meaning “self,” and “telos,” meaning “goal”). “This is a form of curiosity in which individuals create their own goals, like children when they invent games with arbitrary rules,” explains Pierre-Yves Oudeyer. “These two theories open up many new avenues for understanding human learning.”

The second area explored by Flowers is transposing the principles of understanding curiosity to artificial intelligence systems equipped with a certain form of autotelic curiosity, which enables them to learn from their experience and improve by trying to achieve self-generated goals. “Outside our small community of ‘developmental robotics’—as we called it 10 years ago—curiosity algorithms were unknown,” says Pierre-Yves Oudeyer. “We were among the first to test them on machines. Today, things have changed a lot, and curiosity is commonly used in AI and machine learning.”

Until a few years ago, experiments with robots were limited to very basic tasks such as moving objects and stacking them. The team therefore sought to model more creative, abstract forms of objectives by exploiting language. Language is not only used for communication, but also becomes a cognitive tool for inventing goals in a “curious agent,” for example in a video game environment. This agent uses language to create new and abstract objectives.

In 2022, Flowers integrated autotelic curiosity algorithms with foundational models (large language models, LLMs). This made it possible to develop curious and autotelic generative AI agents capable of generating their own increasingly complex goals and training on them, enabling open-ended learning in which they can self-improve. Two key pieces of work in achieving this result were the GLAM system (International Conference on Machine Learning ICML 2023), which transforms LLMs into agents that learn to solve tasks in interactive environments, and the Magellan system (ICML 2024), which gives these generative AI agents the ability to metacognitively predict their own learning progress, enabling them to navigate very large goal spaces efficiently.

The team is now working on applying this approach to enable generative AI agents to self-improve in the areas of code and formal proof in mathematics, in collaboration with the Inria LLM4Code challenge.

The team's third area of focus: applications of curiosity algorithms in the fields of research and education. “Scientists have something in common with children when it comes to curiosity: they seek to explore knowledge and make discoveries, while having limited resources. They therefore need to be efficient,” says Pierre-Yves Oudeyer. "For example, we used curiosity algorithms to help biologists study complex networks, such as those involved in gene regulation. This led us to discover new behaviors that were previously unknown in biology." 

Curiosity algorithms are also particularly useful in the field of artificial life, which studies theoretical biology questions about the origin of life or the formation of organized patterns, such as those on the skin of zebras or leopards. “In a recent article published in Science Advances, where autotelic curiosity algorithms are used to explore self-organization in cellular automata, we showed the possibility of creating protocells capable of moving, avoiding obstacles, or regenerating themselves, similar to what is found in the simplest living beings.”

In the field of education, the Flowers team has used its research to develop teaching tools that stimulate motivation and learning at all ages. Based on advances in understanding human curiosity, particularly the theory of learning progress, researchers have developed teaching methods and AI methods that identify which exercises will help each student progress the most at a given time. 

This collaboration has resulted in two tools: Adaptativ’Math for learning mathematics in primary school, winner of the P2IA Innovation and Artificial Intelligence Partnership, and M.I.A. Seconde, personalized support software for French and mathematics. Both tools are now being rolled out in all primary schools and 10th grade classes in France by the Ministry of Education. And the team does not intend to stop there: it is exploring other areas of application in education, such as training children to ask curious questions and develop their metacognition, training adults, particularly the elderly, to improve their attention span, learning in immersive environments and links with cognitive load, and designing tools to better help students with cognitive disorders.

“The subject of curiosity is inexhaustible,” summarizes Pierre-Yves Oudeyer. “The renewal of our Inria project team allows us to continue our explorations, both fundamental and applied.” This new stage also opens up a new field for Flowers: curiosity on a collective scale, not just on an individual level. So, how can a group of curious individuals innovate? What role does the cultural and social environment play in the emergence of curiosity? 

“We began this study during the summer, freeing ourselves from the Western human models that had guided our work until then,” says the head of Flowers. "One of our doctoral students traveled to Africa to test our learning theories on two populations living in the heart of the Congolese forest. The goal was to determine how cultural differences between sedentary villagers on one hand and a nomadic tribe of hunter-gatherers on the other affect the way curiosity works. The results are currently being analyzed. We can't wait to see them!"

• Open-endedness, curiosity-driven AI and automated search in artificial life, presentation by Pierre-Yves Oudeyer at IMOL 2024, January 20, 2025.
• Curiosity-driven Autotelic AI Agents that Use and Ground Large Language Models, presentation by Pierre-Yves Oudeyer (2024)
• Curiosity and meta-cognition: cognitive sciences, AI, and education, video by Pierre-Yves Oudeyer, presentation given to the Scientific Council of National Education (CSEN), March 20, 2024.
• Curious machines for curious humans, video by Hélène Sauzéon and Clément Moulin-Frier, 4/6/2023.
• Adaptiv'Math, the adaptive digital companion for learning mathematics in cycle 2 (video), EvidenceB, 12/16/2021.