2025Activity report​​Project-TeamFLOWERS

3.2.1‌ Curiosity, meta-cognitionand agency across‌​‌ the lifespan.

The Learning​​ Progress hypothesis, as well​​​‌ as other theories of‌ curiosity-driven learning, all assume‌​‌ meta-cognitive competencies (e.g. ability​​ to evaluate one’s own​​​‌ uncertainty, knowledge gaps or‌ learning progress) as well‌​‌ as forms of agency.​​ However, experimental studies of​​​‌ human curiosity have so‌ far mostly overlooked studying‌​‌ the influence of meta-cognition​​ and agency, let alone​​​‌ simply measure them together‌ with various dimensions of‌​‌ curiosity 132. Another​​ major limit of current​​​‌ models and experimental studies‌ of human curiosity has‌​‌ been that they have​​ not studied how curiosity​​​‌ develops across the lifespan.‌ Actually, the scientific community‌​‌ knows very little on​​ how various forms of​​​‌ curiosity change across childhood,‌ adolescence, and up to‌​‌ ageing populations.

We will​​ aim to address some​​​‌ of these limitations by‌ collaborating with various international‌​‌ groups, including M. Gruber​​ (Univ. Cardiff) and Y.​​​‌ Fandakova (Max Planck Institute‌ for Human Development) with‌​‌ whom we just submitted​​ a major ANR/DFG/ESRC project​​​‌ on this topic. In‌ particular, we propose an‌​‌ interdisciplinary approach to make​​ new breakthroughs in understanding​​​‌ how metacognition contributes to‌ the development of curiosity-based‌​‌ learning, and set the​​ stage for educational interventions​​​‌ that could help children‌ develop their curiosity. Given‌​‌ the links between curiosity​​ and metacognition, and the​​​‌ fact that metacognition continues‌ to improve across childhood‌​‌ and adolescence, we formulate​​ the hypothesis that the​​​‌ efficiency of curiosity-based learning,‌ i.e. the ability to‌​‌ inquire about and prioritise​​ learning of information associated​​​‌ with high curiosity, improves‌ across child and adolescent‌​‌ development.

3.2.2 Experimental paradigms​​ for studying autotelic learning​​​‌ in humans.

Another limit‌ of existing experimental studies‌​‌ of human curiosity, including​​ the ones mentioned above,​​​‌ has been that most‌ of them focused so‌​‌ far on studying how​​ humans prefer exploring one​​​‌ of several pre-existing stimuli‌ or learning activities 163‌​‌. However, as shown​​ in our theoretical and​​​‌ AI work described above,‌ and as argued in‌​‌ complementary arguments from Laura​​ Schulz and Junyi Chu​​​‌ 84, exploration of‌ self-generated goals, including arbitrary‌​‌ goals or games, may​​ be key in accounting​​​‌ for human development, and‌ further in accounting for‌​‌ human innovation and cultural​​ evolution 85. Only​​​‌ very few exploratory experimental‌ protocols have started to‌​‌ be investigated in the​​ literature 93: we​​​‌ aim to further develop‌ this form of experimental‌​‌ protocol, informed by predictions​​ made by our theoretical​​​‌ models, in collaboration with‌ researchers such as G.‌​‌ Molinaro and A. Collins​​ (Univ. Berkeley, both in​​​‌ a 6 months research‌ visit at Inria Flowers‌​‌ and Mnemosyne in 2024)​​ J. Chu (Harvard Univ,​​​‌ US), L. Rat-Fischer (Univ.‌ Nanterre) and A. Ruggeri‌​‌ (TU Munich).

3.2.3 Links​​ between curiosity and creativity​​​‌ for autotelic learning in‌ children:

The ability to‌​‌ imagine abstract and new​​ goals is essential for​​​‌ creative discovery and open-ended‌ learning throughout life. Children‌​‌ achieve this by using​​​‌ the compositionality of language​ as a tool to​‌ imagine situations they have​​ never experienced before, targeting​​​‌ them as goals during​ play 147, 168​‌. Echoing the IMAGINE​​ architecture 88, an​​​‌ intrinsically motivated deep reinforcement​ learning architecture modelling compositional​‌ imagination (the creation of​​ new linguistic associations for​​​‌ new goals), we aim​ to investigate the links​‌ between curiosity and creativity​​ in humans, focusing on​​​‌ the metacognitive role of​ language in guiding autonomous​‌ learning behaviours. Although the​​ nature of the links​​​‌ between curiosity and creativity​ is currently not well​‌ defined, a recent meta-analysis​​ shows that higher levels​​​‌ of curiosity are significantly​ associated with higher levels​‌ of creativity 156.​​ Divergent thinking mechanisms are​​​‌ said to be the​ cognitive resource common to​‌ both skills 104,​​ 117, and some​​​‌ authors even identify curiosity​ as a facilitator, a​‌ trigger for creativity 106​​: high curiosity states​​​‌ induce better ideation and​ greater idea associations conducive​‌ to problem solving. Also,​​ both creativity and curiosity​​​‌ are governed by metacognitive​ processes of self-regulation of​‌ learning 151 enabling the​​ identification of information gaps,​​​‌ problem situations or uncertainties,​ the generation of ideas,​‌ paths to resolution, and​​ monitoring and evaluating the​​​‌ value of ideas as​ creative output or as​‌ majoring knowledge 109,​​ 112. We aim​​​‌ to investigate developmental differences​ on curiosity-based learning and​‌ problem-solving tasks while studying​​ their relationships and their​​​‌ dependency to intrapersonal factors​ (especially metacognitive skills and​‌ personality dimensions such as​​ epistemic curiosity, creativity or​​​‌ intellectual humility traits) in​ late childhood (from 6​‌ to 11 yo). To​​ achieve experiments needed to​​​‌ address these topics, we​ will leverage an educational​‌ Léa-Ifé collaboration network established​​ with 10 primary schools​​​‌ around Bordeaux. As a​ whole, in this part​‌ of the project, we​​ aim to demonstrate that​​​‌ curiosity as a process​ to seek knowledge in​‌ the face of self-generated​​ goals of knowledge gaps,​​​‌ or as a metacognitive​ feeling 103, leads​‌ to better initiation of​​ the creative process.

3.2.4​​​‌ Curiosity to learn about​ others and social interaction:​‌

Social curiosity is defined​​ as the desire to​​​‌ acquire knowledge about others​ in society, encompassing an​‌ interest in their emotions,​​ thoughts, and behaviours. This​​​‌ type of curiosity can​ be divided into two​‌ forms 146: 1)​​ empathetic curiosity (the desire​​​‌ to acquire knowledge about​ others), and 2) relational​‌ curiosity (the desire to​​ interact with others). Like​​​‌ other types of curiosity,​ social curiosity motivates people​‌ to engage in exploratory​​ behaviours directed toward the​​​‌ social world, seeking novel​ information about how people​‌ think, behave, and feel.​​ 110 proposes three functions​​​‌ of social curiosity: 1)​ acquiring information useful for​‌ learning and development, 2)​​ establishing interpersonal relationships and​​​‌ increasing a sense of​ social belonging, and 3)​‌ controlling the social world​​ by making it more​​​‌ predictable and manageable. Thus,​ social curiosity enhances social​‌ functioning and has been​​ linked to improved social​​​‌ behaviour adaptation, the ability​ to establish and maintain​‌ social relationships, and better​​ social judgement abilities 110​​. Recently, another distinction​​​‌ has been proposed in‌ social curiosity 114:‌​‌ 1) overt social curiosity,​​ an explicit interest in​​​‌ understanding other people, which‌ motivates direct communication with‌​‌ others; 2) covert social​​ curiosity, an “hidden” interest​​​‌ that motivates more indirect‌ and furtive behaviours to‌​‌ understand others, such as​​ discreetly observing people, listening​​​‌ to others’ conversations, and‌ reading tabloids and human-interest‌​‌ stories. Covert curiosity is​​ often associated with negative​​​‌ outcomes like gossiping or‌ spying 114, but‌​‌ it can also drive​​ the understanding of the​​​‌ social world through observation‌ and finally motive interactions‌​‌ with others 105.​​ On the other hand,​​​‌ overt social curiosity has‌ been linked with open-mindedness,‌​‌ extraversion, and sociability 114​​, and was associated​​​‌ with better job performance‌ 105.

3.2.5 Autotelic‌​‌ game invention and cultural​​ transmission.

Leveraging the theoretical​​​‌ ideas on the interaction‌ between autotelic learning and‌​‌ cultural evolution as described​​ in the previous section,​​​‌ we also aim to‌ study experimentally these interactions‌​‌ in chains of humans​​ incentivized to transmit to​​​‌ each other games or‌ artefacts of their own‌​‌ intrinsically motivated invention (either​​ physical or video games,​​​‌ e.g. using experimental setups‌ like 93). We‌​‌ aim to design new​​ experimental protocols and run​​​‌ them both in various‌ age ranges in European‌​‌ populations, as well as​​ in populations in non​​​‌ western culture leveraging associated‌ collaborations with Maxime Derex‌​‌ at IAST, Toulouse, Sheina​​ Lew-Levy at Durham University,​​​‌ and Sarah Pope-Caldwell at‌ Georgia State University.

3.3.1 Language-Augmented​​​‌ Autotelic Agents with Foundational‌ Models

We will develop‌​‌ architectures where LLMs function​​ as cognitive tools for​​​‌ autotelic RL agents across‌ five dimensions: (1) LLM-based‌​‌ agents with environmental alignment—extending​​ our work on grounding​​​‌ LLMs through online RL‌ 79 where LLMs generate‌​‌ goals, evaluate achievement, relabel​​ experiences, and provide natural​​​‌ language interfaces. We will‌ extend goal generation to‌​‌ creative, time-extended, and learning-oriented​​ goals including self-generated causal​​​‌ questions and hypotheses 101‌, while correcting hallucinations‌​‌ through incremental LLM updates​​ via environment interaction. (2)​​​‌ Multimodal grounding and social‌ environments—extending our SocialAI School‌​‌ framework 118 to incorporate​​ theory of mind, joint​​​‌ intentionality, and social norms,‌ investigating whether social curiosity‌​‌ can drive efficient acquisition​​ of complex social skills.​​​‌ (3) Real-time human-in-the-loop learning—enabling‌ agents to interpret instructions,‌​‌ respond to feedback, explain​​ exploration processes, and adapt​​​‌ to user preferences for‌ education and discovery applications.‌​‌ (4) Learning to use​​ cognitive tools—agents will learn​​​‌ when to invoke APIs,‌ generate/execute code, query knowledge‌​‌ bases, or request human​​ assistance, including chain-of-thought and​​​‌ self-reflection mechanisms. (5) Metacognitive‌ curriculum learning with coordinated‌​‌ interestingness measures—leveraging our MAGELLAN​​ architecture 47 which enables​​​‌ LLM agents to learn‌ metacognitive predictions of their‌​‌ own competence and learning​​ progress across large language-defined​​​‌ goal spaces. By capturing‌ semantic relationships between goals,‌​‌ MAGELLAN enables sample-efficient progress​​ estimation and dynamic adaptation​​​‌ to evolving goal spaces.‌ We will extend this‌​‌ to develop meta-diversity search​​ algorithms 97 leveraging LLMs​​​‌ to generate novel conceptual‌ dimensions, enabling exploration across‌​‌ objective (learning progress, novelty)​​​‌ and subjective, culturally-contextualized criteria​ 158. Long-term objectives​‌ will include studying how​​ agents pursue goals across​​​‌ extended timescales using cultural​ artifacts for long-term planning.​‌

3.3.2 Program Synthesis for​​ Abstract and Verifiable Intelligence​​​‌

We will explore autotelic​ learning in formal language​‌ spaces where goals and​​ policies are represented as​​​‌ programs. Our ACES architecture​ demonstrates autotelic LLMs self-improving​‌ coding skills by iteratively​​ generating diverse problems and​​​‌ solutions using code interpreters​ 150. This addresses​‌ three limitations: environments are​​ not truly open-ended (code​​​‌ is), LLMs lack grounding​ (interpreters provide it), and​‌ code LLMs struggle beyond​​ training distributions (autotelic learning​​​‌ enables self-improvement). We will​ train small models with​‌ advanced coding capabilities and​​ extend to mathematical problem​​​‌ invention and theorem proving.​ Within Inria LLM4Code, we​‌ will develop autotelic LLMs​​ interacting with proof assistants​​​‌ like Coq and Lean​ 157. This approach​‌ provides compact interpretable representations,​​ formal verification, and compositional​​​‌ generalization. Progress on benchmarks​ like ARC 83 will​‌ validate these methods.

3.3.3​​ Curiosity in Cultural Evolution,​​​‌ Collective Intelligence, and AI​ Science Teams

Understanding how​‌ groups coordinate curiosity-driven exploration​​ and self-organize collective intelligence​​​‌ represents both a fundamental​ scientific question and a​‌ path toward transformative applications.​​ This research direction addresses​​​‌ several interconnected challenges. First,​ how individual curiosity combines​‌ with social transmission and​​ collective innovation 111,​​​‌ 95 is still poorly​ understood and modeled. Second,​‌ as generative AI increasingly​​ participates in human cultural​​​‌ production, understanding cultural evolution​ in hybrid human-AI groups​‌ becomes essential for anticipating​​ societal impacts. Third, many​​​‌ scientific and creative challenges​ require coordinated teams leveraging​‌ complementary expertise and perspectives—motivating​​ our vision of autotelic​​​‌ AI science teams collaborating​ with human researchers.

Near-term​‌ work will investigate coordination​​ when self-generated goals conflict​​​‌ in shared environments. Some​ goals require collaboration with​‌ agents possessing complementary skills—agents​​ must negotiate joint goals​​​‌ serving individual curiosity. We​ will study how network​‌ topology influences innovation dynamics​​ 137, how agents​​​‌ develop communication protocols 75​, and whether groups​‌ display curiosity at collective​​ levels—pursuing structured exploration maximizing​​​‌ diversity of learned goals​ through simple individual-level mechanisms.​‌ The increasing role of​​ generative AI in cultural​​​‌ production necessitates understanding these​ dynamics in hybrid settings—the​‌ emerging field of "Machine​​ Culture" 77. We​​​‌ will systematically investigate how​ interaction protocols, social structures,​‌ and model capabilities shape​​ cultural dynamics in LLM​​​‌ populations and mixed human-AI​ groups (collaboration with Derex,​‌ IAST).

3.4.1 Training Curiosity and​ Metacognition Across the Lifespan​‌

Addressing 21st century educational​​ challenges—inclusive education, cross-disciplinary skills​​​‌ (attention, curiosity, learning to​ learn), and digital transformation—requires​‌ technologies fostering curiosity and​​ metacognition. Our ZPDES algorithm​​​‌ personalizes curricula by maximizing​ learning progress 87.​‌ RCTs with >1,000 children​​ demonstrated enhanced learning efficiency​​​‌ and motivation versus expert-designed​ curricula. We will extend​‌ ZPDES to new domains​​ (attention training, language learning)​​​‌ and populations: aging adults,​ neurodiverse learners, professional contexts​‌ (sports, gaming). To address​​ ZPDES's requirement for expert-formatted​​​‌ content, we will leverage​ generative AI to automate​‌ exercise generation from textbooks,​​ training smaller LLMs for​​ lightweight systems avoiding foreign-hosted​​​‌ dependencies. We will scale‌ up metacognitive skills and‌​‌ curious question training 69​​, studying transfer to​​​‌ creativity and including pioneering‌ work using GPT-3 conversational‌​‌ agents in real classrooms​​ 70.

A distinctive​​​‌ objective involves metacognitive empowerment:‌ developing interventions helping children‌​‌ understand their own learning​​ progress to self-generate curricula​​​‌ independently—addressing limited technology contexts‌ while fostering autonomy. Long-term‌​‌ work will include teacher​​ training programs embedding curiosity-fostering​​​‌ practices (Peterson, 2020) and‌ unplugged activity versions. Partnerships‌​‌ with educational institutions (Académie​​ de Bordeaux), industry (EvidenceB,​​​‌ Ubisoft), and NGOs (France‌ IOI) will enable deployment‌​‌ and policy influence.

3.4.2​​ Assisted Scientific Discovery with​​​‌ Autotelic Exploration

Scientists studying‌ complex systems face challenges‌​‌ mapping behavioral spaces when​​ lacking models and representations​​​‌ with scarce experimental resources.‌ Autotelic algorithms offer sample-efficient‌​‌ diverse behavior discovery while​​ remaining steerable through natural​​​‌ language. Proof-of-concept work efficiently‌ maps spaces in cellular‌​‌ automata 98 and gene​​ regulatory networks (99​​​‌, with Levin, Harvard),‌ independently adopted by physics‌​‌ researchers (U. Washington). We​​ will maximize diversity through​​​‌ meta-diversity search and leverage‌ language/multimodal models for abstract‌​‌ guidance.

Near-term work will​​ leverage artificial life as​​​‌ testbeds, studying origins of‌ autopoietic systems and self-organizing‌​‌ evolutionary processes in cellular​​ automata 34, 149​​​‌ (2023 Best Paper ALife).‌ Long-term objectives will transition‌​‌ to real-world systems through​​ collaborations: Levin (synthetic biology),​​​‌ Murugan (soft condensed matter,‌ U. Chicago), Aymonier (chemistry,‌​‌ ICMCB Bordeaux), with applications​​ to power networks, neuromuscular​​​‌ models, and artistic domains.‌ We will explore autotelic‌​‌ algorithms for mathematical problem/proof​​ exploration within LLM4Code. Success​​​‌ will establish autotelic exploration‌ as methodology for materials‌​‌ science, systems biology, and​​ mathematical discovery, with industrial​​​‌ translation potential (e.g. Solvay/Syensqo).‌

7.1.1 SocialAI​‌

• Name:
  SocialAI: Benchmarking Socio-Cognitive​​ Abilities in Deep Reinforcement​​ Learning Agents
• Keywords:
  Artificial​​​‌ intelligence, Deep learning, Reinforcement‌ learning, Large Language Models‌​‌
• Functional Description:

  Source code​​ for the paper https://arxiv.org/abs/2107.00956.​​​‌

  A suite of environments‌ for testing socio-cognitive abilities‌​‌ of artificial agents. Environments​​ can be used in​​​‌ the multimodal setting (suitable‌ for RL agents) and‌​‌ in the pure text​​ setting (suitable for Large​​​‌ Language Model-based agents). Also‌ contains RL and LLM‌​‌ baselines.

• URL:
  https://­gitlab.­inria.­fr/­gkovac/­act-and-speak
• Contact:​​
  Grgur Kovac

7.1.2 AutoDisc​​​‌

• Keyword:
  Complex Systems
• Functional‌ Description:
  AutoDisc is a‌​‌ software built for automated​​ scientific discoveries in complex​​​‌ systems (e.g. self-organizing systems).‌ It can be used‌​‌ as a tool to​​ experiment automated discovery of​​​‌ various systems using exploration‌ algorithms (e.g. curiosity-driven). Our‌​‌ software is fully Open​​ Source and allows user​​​‌ to add their own‌ systems, exploration algorithms or‌​‌ visualization methods.
• URL:
  https://­gitlab.­inria.­fr/­cromac/­AutomatedDiscoveryTool​​
• Contact:
  Clément Romac

7.1.3​​​‌ ADTool

• Keywords:
  Machine learning,‌ Python, Cellular automaton, Physical‌​‌ simulation, Pattern discovery, Exploration​​
• Functional Description:

  ADTool is​​​‌ a versatile and open-source‌ Python framework designed to‌​‌ explore complex parametric systems​​ using IMGEP algorithms (Intrinsic​​​‌ Motivation for Goal Exploration‌ Processes) as described in‌​‌ https://arxiv.org/pdf/1708.02190. This curiosity-driven approach​​ enables automatic exploration and​​​‌ the discovery of new‌ behaviors across a wide‌​‌ range of domains, offering​​ a novel way to​​​‌ study complex systems.

  With‌ ADTool, users can explore‌​‌ cellular automata such as​​ Lenia, Particle Lenia, and​​​‌ Flowlenia to uncover patterns‌ and emergent behaviors. Its‌​‌ capabilities extend to drug​​ discovery, exploring chemical spaces​​​‌ to identify promising protein-ligand‌ affinity profiles. The framework‌​‌ also ventures into physics,​​ with applications such as​​​‌ searching for trajectories in‌ the N-body problem, simulating‌​‌ the Kuramoto model, exploring​​ the Gray-Scott reaction-diffusion system,​​​‌ and studying hypergraph rewriting‌ systems for Wolfram physics.‌​‌ In digital art, ADTool​​ fosters creativity by exploring​​​‌ processes like subtractive sound‌ synthesis and other artistic‌​‌ methods.

  The framework is​​ designed to be flexible​​​‌ and extensible, allowing users‌ to define their own‌​‌ systems and integrate custom​​ exploration strategies. It includes​​​‌ mechanisms for saving discoveries‌ to disk, making it‌​‌ easier to resume experiments​​ or share results with​​​‌ collaborators. Additionally, an integrated‌ visualization tool provides a‌​‌ user-friendly interface to track​​ exploration progress, enhancing the​​​‌ understanding and analysis of‌ results.

  The scientific foundation‌​‌ of ADTool lies in​​ "curiosity-search" algorithms, which autonomously​​​‌ explore behavioral spaces to‌ identify interesting phenomena without‌​‌ predefined objectives. These algorithms,​​ initially developed for robotic​​​‌ learning, are now applied‌ to the study of‌​‌ emergent behaviors in various​​ systems.

  Whether you are​​​‌ a physicist, chemist, biologist,‌ or digital artist, ADTool‌​‌ can help you explore​​ and understand complex systems.​​​‌

  Reproducibility is guarantied with‌ a predifined Python environment‌​‌ and experiments can be​​ launched with a simple​​​‌ command line: python3 run.py‌ –config_file examples/grayscott/gray_scott.json

• Contact:
  Zacharie‌​‌ Bugaud

7.1.4 Kids Ask​​

• Keywords:
  Human Computer Interaction,​​​‌ Cognitive sciences
• Functional Description:‌
  Kids Ask is a‌​‌ web-based educational platform that​​ involves an interaction between​​​‌ a child and a‌ conversational agent. The platform‌​‌ is designed to teach​​ children how to generate​​​‌ curiosity-based questions and use‌ them in their learning‌​‌ in order to gain​​​‌ new knowledge in an​ autonomous way.
• URL:
  https://­github.­com/­RaniaAbdelghani/­KidAsk​‌
• Contact:
  Rania Abdelghani

7.1.5​​ ToGather

• Keywords:
  Education, Handicap,​​​‌ Environment perception
• Scientific Description:​
  With participatory design methods,​‌ we have designed an​​ interactive website application for​​​‌ educational purposes. This application​ aims to provide interactive​‌ services with continuously updated​​ content for the stakeholders​​​‌ of school inclusion of​ children with specific educational​‌ needs.
• Functional Description:
  Website​​ gathering information on middle​​​‌ school students with neurodevelopmental​ disorders. Authentication is required​‌ to access the site's​​ content. Each user can​​​‌ only access the student​ file(s) of the young​‌ person(s) they are accompanying.​​ A student file contains​​​‌ 6 tabs, in which​ each type of user​‌ can add, edit or​​ delete information: 1. Profile:​​​‌ to quickly get to​ know the student 2.​‌ Skills: evaluation at a​​ given moment and evolution​​​‌ over time 3. Compendium​ of tips: includes psycho-educational​‌ tips 4. Meetings: manager​​ and reports 5. News:​​​‌ share information over time​ 6. Contacts: contact information​‌ for stakeholders The student​​ only has the right​​​‌ to view information about​ him/her.
• Publication:
  hal-03436355
• Contact:​‌
  Cécile Mazon
• Participant:
  4​​ anonymous participants

7.1.6 mc_training​​​‌

• Name:
  Platform for metacognitive​ training
• Keywords:
  Human Computer​‌ Interaction, Education
• Functional Description:​​

  This is a web​​​‌ platform for children between​ 9 and 11 years​‌ old, designed to help​​ children practice 4 metacognitive​​​‌ skills that are thought​ to be involved in​‌ curiosity-driven learning: - the​​ ability to identify uncertainties​​​‌ - the ability to​ generate informed hypotheses -​‌ the ability to ask​​ questions - the ability​​​‌ to evaluate the value​ of a preconceived inference.​‌

  Children work on a​​ reading-comprehension tasks and, for​​​‌ each of these skills,​ the platform offers help​‌ through a "conversation" with​​ conversational agents that give​​​‌ instructions to perform the​ task, with respect to​‌ every skill, and can​​ give suggestions if the​​​‌ child asks for it.​

• Contact:
  Rania Abdelghani

7.1.7​‌ Evolution of adaptation mechanisms​​ in complex environments

• Name:​​​‌
  Plasticity and evolvability under​ environmental variability: the joint​‌ role of fitness-based selection​​ and niche-limited competition
• Keywords:​​​‌
  Evolution, Ecology, Dynamic adaptation​
• Functional Description:

  This is​‌ the code accompannying our​​ paper Plasticity and evolvability​​​‌ under environmental variability: the​ joint role of fitness-based​‌ selection and niche-limited competition"​​ which is to be​​​‌ presented at the Gecco​ 2022 conference.

  In this​‌ work we have studied​​ the evolution of a​​​‌ population of agents in​ a world where the​‌ fitness landscape changes with​​ generations based on climate​​​‌ function and a latitudinal​ model that divides the​‌ world in different niches.​​ We have implemented different​​​‌ selection mechanisms (fitness-based selection​ and niche-limited competition).

  The​‌ world is divided into​​ niches that correspond to​​​‌ different latitudes and whose​ state evolves based on​‌ a common climate function.​​

  We model the plasticity​​​‌ of an individual using​ tolerance curves originally developed​‌ in ecology. Plasticity curves​​ have the form of​​​‌ a Gaussian the capture​ the benefits and costs​‌ of plasticity when comparing​​ a specialist (left) with​​​‌ a generalist (right) agent.​

  The repo contains the​‌ following main elements :​​

  folder source contains the​​ main functionality for running​​​‌ a simulation scripts/run/reproduce_gecco.py can‌ be used to rerun‌​‌ all simulations in the​​ paper scripts/evaluate contains scripts​​​‌ for reproducing figures. reproduce_figures.py‌ will produce all figures‌​‌ (provided you have already​​ run scripts/run/reproduce_gecco.py to generate​​​‌ the data) folder projects‌ contains data generated from‌​‌ running a simulation How​​ to run To install​​​‌ all package dependencies you‌ can create a conda‌​‌ environment as:

  conda env​​ create -f environment.yml

  All​​​‌ script executions need to‌ be run from folder‌​‌ source. Once there, you​​ can use simulate.py, the​​​‌ main interface of the‌ codebase to run a‌​‌ simulation, For example:

  python​​ simulate.py –project test_stable –env_type​​​‌ stable –num_gens 300 –capacity‌ 1000 –num_niches 10 –trials‌​‌ 10 –selection_type NF –climate_mean_init​​ 2

  will run a​​​‌ simulation with an environment‌ with a climate function‌​‌ whose state is constantly​​ 2 consisting of 100​​​‌ niches for 300 generations‌ and 10 independent trials.‌​‌ The maximum population size​​ will be 1000*2 and​​​‌ selection will be fitness-based‌ (higher fitness means higher‌​‌ chances of reproduction) and​​ niche limited (individuals reproduce​​​‌ independently in each niche‌ and compete only within‌​‌ a niche),

  You can​​ also take a look​​​‌ at scripts/run/reproduce_gecco.py to see‌ which flags were used‌​‌ for the simulations presented​​ in the paper.

  Running​​​‌ all simulations requires some‌ days. You can instead‌​‌ download the data produced​​ by running scripts/run/reproduce_gecco.py from​​​‌ this google folder and‌ unzip them under the‌​‌ projects directory.

• URL:
  https://­github.­com/­eleninisioti/­ClimateAndLearning​​
• Contact:
  Eleni Nisioti

7.1.8​​​‌ SAPIENS

• Name:
  SAPIENS: Structuring‌ multi-Agent toPology for Innovation‌​‌ through ExperieNce Sharing
• Keywords:​​
  Reinforcement learning, Multi-agent
• Functional​​​‌ Description:

  SAPIENS is a‌ reinforcement learning algorithm where‌​‌ multiple off-policy agents solve​​ the same task in​​​‌ parallel and exchange experiences‌ on the go. The‌​‌ group is characterized by​​ its topology, a graph​​​‌ that determines who communicates‌ with whom.

  All agents‌​‌ are DQNs and exchange​​ experiences have the form​​​‌ of transitions from their‌ replay buffers.

  Using SAPIENS‌​‌ we can define groups​​ of agents that are​​​‌ connected with others based‌ on a a) fully-connected‌​‌ topology b) small-world topology​​ c) ring topology or​​​‌ d) dynamic topology.

  Install‌ required packages You can‌​‌ install all required python​​ packages by creating a​​​‌ new conda environment containing‌ the packages in environment.yml:‌​‌

  conda env create -f​​ environment.yml

  And then activating​​​‌ the environment:

  conda activate‌ sapiens

  Example usages Under‌​‌ notebooks there is a​​ Jupyter notebook that will​​​‌ guide you through setting‌ up simulations with a‌​‌ fully-connected and a dynamic​​ social network structure for​​​‌ solving Wordcraft tasks. It‌ also explains how you‌​‌ can access visualizations of​​ the metrics produced during​​​‌ th$

  Reproducing the paper‌ results Scripts under the‌​‌ scripts directory are useful​​ for reproducing results and​​​‌ figures appearing in the‌ paper.

  With scripts/reproduce_runs.py you‌​‌ can run all simulations​​ presented in the paper​​​‌ from scratch.

  This file‌ is useful for looking‌​‌ at how the experiments​​ were configured but better​​​‌ avoid running it: simulations‌ will run locally and‌​‌ sequentially and will take​​ months to complete.

  Instead,​​​‌ you can access the‌ data files output by‌​‌ simulations on this online​​​‌ repo.

  Download this zip​ file and uncompress it​‌ under the projects directory.​​ This should create a​​​‌ projects/paper_done sub-directory.

  You can​ now reproduce all visualization​‌ presented in the paper.​​ Run:

  python scripts/reproduce_visuals.py

  This​​​‌ will save some general​ plots under visuals, while​‌ project-specific plots are saved​​ under the corresponding project​​​‌ in projects/paper_done

• URL:
  https://­github.­com/­eleninisioti/­SAPIENS​
• Contact:
  Eleni Nisioti

7.1.9​‌ architect-builder-abig

• Name:
  Architect-Builder Iterated​​ Guiding
• Keyword:
  Artificial intelligence​​​‌
• Functional Description:

  Codebase for​ the paper Learning to​‌ guide and to be​​ guided in the Architect-Builder​​​‌ Problem

  ABIG stands for​ Architect-Builder Iterated Guiding and​‌ is an algorithmic solution​​ to the Architect-Builder Problem.​​​‌ The algorithm leverages a​ learned model of the​‌ builder to guide it​​ while the builder uses​​​‌ self-imitation learning to reinforce​ its guided behavior.

• URL:​‌
  https://­github.­com/­flowersteam/­architect-builder-abig
• Contact:
  Tristan Karch​​

7.1.10 EAGER

• Name:
  Exploit​​​‌ question-Answering Grounding for effective​ Exploration in language-conditioned Reinforcement​‌ learning
• Keywords:
  Reinforcement learning,​​ Language, Question Generation Question​​​‌ Answering, Reward shaping
• Functional​ Description:
  A novel QG/QA​‌ framework for RL called​​ EAGER In EAGER, an​​​‌ agent reuses the initial​ language goal sentence to​‌ generate a set of​​ questions (QG): each of​​​‌ these self-generated questions defines​ an auxiliary objective. Here,​‌ generating a question consists​​ in masking a word​​​‌ of the initial language​ goal. Then the agent​‌ tries to answer these​​ questions (guess the missing​​​‌ word) only by observing​ its trajectory so far.​‌ When it manages to​​ answer a question correctly​​​‌ (QA) it obtains an​ intrinsic reward proportional to​‌ its confidence in the​​ answer. The QA module​​​‌ is trained using a​ set of successful example​‌ trajectories. If the agent​​ follows a path too​​​‌ different from correct ones​ at some point in​‌ its trajectory, the QA​​ module will not answer​​​‌ the question correctly, resulting​ in zero intrinsic reward.​‌ The sum of all​​ the intrinsic rewards measures​​​‌ the quality of a​ trajectory in relation to​‌ the given goal. In​​ other words, maximizing this​​​‌ intrinsic reward incentivizes the​ agent to produce behaviour​‌ that unambiguously explains various​​ aspects of the given​​​‌ goal.
• URL:
  https://­github.­com/­flowersteam/­EAGER
• Contact:​
  Thomas Carta

7.1.11 Flow-Lenia​‌

• Name:
  Flow Lenia: Mass​​ conservation for the study​​​‌ of virtual creatures in​ continuous cellular automata
• Keywords:​‌
  Cellular automaton, Self-organization
• Functional​​ Description:

  This repo contains​​​‌ the code to run​ the Flow Lenia system​‌ which is a continuous​​ parametrized cellular automaton with​​​‌ mass conservation. This work​ extends the classic Lenia​‌ system with mass conservation​​ and allows to implement​​​‌ new feature like local​ parameter, environment components etc​‌

  Several declination of the​​ system (1 or several​​​‌ channels etc ) are​ available

  Please refer to​‌ the associated paper for​​ the details of the​​​‌ system

  Implemented in JAX​

• URL:
  https://­github.­com/­erwanplantec/­FlowLenia
• Contact:
  Gautier​‌ Hamon

7.1.12 Kidlearn: money​​ game application

• Functional Description:​​​‌
  The games is instantiated​ in a browser environment​‌ where students are proposed​​ exercises in the form​​​‌ of money/token games (see​ Figure 1). For​‌ an exercise type, one​​ object is presented with​​​‌ a given tagged price​ and the learner has​‌ to choose which combination​​ of bank notes, coins​​ or abstract tokens need​​​‌ to be taken from‌ the wallet to buy‌​‌ the object, with various​​ constraints depending on exercises​​​‌ parameters. The games have‌ been developed using web‌​‌ technologies, HTML5, javascript and​​ Django.

  

  

  

  

• URL:
  https://­flowers.­inria.­fr/­research/­kidlearn/
• Contact:‌​‌
  Benjamin Clement

7.1.13 cognitive-testbattery​​

• Name:
  Cognitive test battery​​​‌ of human attention and‌ memory
• Keywords:
  Open Access,‌​‌ Cognitive sciences
• Scientific Description:​​
  Cognitive test batteries are​​​‌ widely used in diverse‌ research fields, such as‌​‌ cognitive training, cognitive disorder​​ assessment, or brain mechanism​​​‌ understanding. Although they need‌ flexibility according to the‌​‌ objectives of their usage,​​ most of the test​​​‌ batteries are not be‌ available as open-source software‌​‌ and not be tuned​​ by researchers in detail.​​​‌ The present study introduces‌ an open-source cognitive test‌​‌ battery to assess attention​​ and memory, using a​​​‌ javascript library, p5.js. Because‌ of the ubiquitous nature‌​‌ of dynamic attention in​​ our daily lives, it​​​‌ is crucial to have‌ tools for its assessment‌​‌ or training. For that​​ purpose, our test battery​​​‌ includes seven cognitive tasks‌ (multiple-objects tracking, enumeration, go/no-go,‌​‌ load-induced blindness, task-switching, working​​ memory, and memorability), common​​​‌ in cognitive science literature.‌ By using the test‌​‌ battery, we conducted an​​ online experiment to collect​​​‌ the benchmark data. Results‌ conducted on two separate‌​‌ days showed the high​​ cross-day reliability. Specifically, the​​​‌ task performance did not‌ largely change with the‌​‌ different days. Besides, our​​ test battery captures diverse​​​‌ individual differences and can‌ evaluate them based on‌​‌ the cognitive factors extracted​​ from latent factor analysis.​​​‌ Since we share our‌ source code as open-source‌​‌ software, users can expand​​ and manipulate experimental conditions​​​‌ flexibly. Our test battery‌ is also flexible in‌​‌ terms of the experimental​​ environment, i.e., it is​​​‌ possible to experiment either‌ online or in a‌​‌ laboratory environment.
• Functional Description:​​
  The evaluation battery consists​​​‌ of 6 cognitive activities‌ (serious games: multi-object tracking,‌​‌ enumeration, go/no-go, Corsi, load-induced​​ blindness, taskswitching, memorability). Easily​​​‌ deployable as a web‌ application, it can be‌​‌ re-used and modified for​​ new experiments. The tool​​​‌ is documented in order‌ to facilitate the deployment‌​‌ and the analysis of​​ results.
• URL:
  https://­github.­com/­flowersteam/­cognitive-testbattery
• Publication:​​​‌
  hal-03723887
• Contact:
  Maxime Adolphe‌
• Participant:
  4 anonymous participants‌​‌

7.1.14 Sensorimotor-lenia

• Keywords:
  Cellular​​ automaton, Gradient descent, Curriculum​​​‌ Learning
• Functional Description:
  Source‌ code for the search‌​‌ of sensorimotor agency in​​ cellular automata associated to​​​‌ this blogpost https://developmentalsystems.org/sensorimotor-lenia/. The‌ code allows to find‌​‌ rules in the cellular​​ automata Lenia (through gradient​​​‌ descent, curriculum learning and‌ diversity search) that lead‌​‌ to the self-organization of​​ moving agents robust to​​​‌ perturbation by obstacles.
• URL:‌
  https://­github.­com/­flowersteam/­sensorimotor-lenia-search
• Contact:
  Gautier Hamon‌​‌

7.1.15 Lamorel

• Keywords:
  Large​​​‌ Language Models, Reinforcement learning,​ Distributed computing
• Scientific Description:​‌

  Lamorel allows for seamless​​ scaling of LLMs when​​​‌ using embodied artificial agents​ such as Reinforcement Learning​‌ agents. One can use​​ and modify the LLM​​​‌ in any part of​ such agents (policy, goal​‌ sampler, social peer...). Lamorel​​ is particularly useful when​​​‌ performing large-scale experiments on​ clusters.

  It was already​‌ used in several papers,​​ notably leading to the​​​‌ first paper performing online​ RL on an LLM-based​‌ agent in an embodied​​ environment (Carta et. al,​​​‌ 2023).

• Functional Description:

  Lamorel​ was initially designed to​‌ easily use LLMs in​​ interactive environments. It is​​​‌ especially made for high​ throughput using a distributed​‌ architecture. The philosophy of​​ *Lamorel* is to be​​​‌ very permissive and allow​ as much as possible​‌ usage of LLMs while​​ maintaining scaling: the application​​​‌ should run with 1​ or N LLMs.

  For​‌ this reason, it is​​ not specialised neither in​​​‌ RL nor in particular​ in RLHF. Our examples​‌ illustrate how *Lamorel* can​​ be used for various​​​‌ applications including RLHF-like finetuning.​ However, one must understand​‌ that *Lamorel*'s philosophy means​​ that users must implement​​​‌ themselves what they want​ to do with the​‌ LLM(s).

  This is why​​ we advise users knowing​​​‌ in advance they want​ to do RLHF, especially​‌ without any modification of​​ classic implementations, to use​​​‌ libs specialised in RLHF​ that already come with​‌ RL implementations (e.g. RL4LMs,​​ TRL). On the other​​​‌ hand, users more inclined​ to experiment with implementations​‌ or looking for an​​ LLM lib they can​​​‌ use in different projects​ may prefer Lamorel.

  Here​‌ are Lamorel's key features:​​ 1. Abstracts the use​​​‌ of LLMs (e.g. tonekization,​ batches) into simple calls​‌

  2. Provides a method​​ to compute the log​​​‌ probability of token sequences​ (e.g. action commands) given​‌ a prompt 3. Is​​ made for scaling up​​​‌ your experiments by deploying​ multiple instances of the​‌ LLM and dispatching the​​ computation thanks to a​​​‌ simple configuration file 4.​ Provides access to open-sourced​‌ LLMs from the Hugging​​ Face's hub along with​​​‌ Model Parallelism to use​ multiple GPUs for an​‌ LLM instance 5. Allows​​ one to give their​​​‌ own PyTorch modules to​ compute custom operations (e.g.​‌ to add new heads​​ on top of the​​​‌ LLM) 6. Allows one​ to train the LLM​‌ (or part of it)​​ thanks to a Data​​​‌ Parallelism setup where the​ user provides its own​‌ update method

• URL:
  https://­github.­com/­flowersteam/­lamorel​​
• Publications:
  hal-03970122, hal-04844089​​​‌, hal-04844077
• Contact:
  Clément​ Romac

7.1.16 GLAM

• Name:​‌
  Grounding LAnguage Models
• Keywords:​​
  Large Language Models, Reinforcement​​​‌ learning
• Scientific Description:
  Recent​ works successfully leveraged Large​‌ Language Models' (LLM) abilities​​ to capture abstract knowledge​​​‌ about world's physics to​ solve decision-making problems. Yet,​‌ the alignment between LLMs'​​ knowledge and the environment​​​‌ can be wrong and​ limit functional competence due​‌ to lack of grounding.​​ In this paper, we​​​‌ study an approach (named​ GLAM) to achieve this​‌ alignment through functional grounding:​​ we consider an agent​​​‌ using an LLM as​ a policy that is​‌ progressively updated as the​​ agent interacts with the​​ environment, leveraging online Reinforcement​​​‌ Learning to improve its‌ performance to solve goals.‌​‌ Using an interactive textual​​ environment designed to study​​​‌ higher-level forms of functional‌ grounding, and a set‌​‌ of spatial and navigation​​ tasks, we study several​​​‌ scientific questions: 1) Can‌ LLMs boost sample efficiency‌​‌ for online learning of​​ various RL tasks? 2)​​​‌ How can it boost‌ different forms of generalization?‌​‌ 3) What is the​​ impact of online learning?​​​‌ We study these questions‌ by functionally grounding several‌​‌ variants (size, architecture) of​​ FLAN-T5.
• Functional Description:
  GLAM​​​‌ is a new approach‌ to achieve alignment between‌​‌ a Large Language Model​​ (LLM) and a considered​​​‌ environment/world through functional grounding:‌ we consider an agent‌​‌ using an LLM as​​ a policy that is​​​‌ progressively updated as the‌ agent interacts with the‌​‌ environment, leveraging online Reinforcement​​ Learning to improve its​​​‌ performance to solve goals.‌
• URL:
  https://­github.­com/­flowersteam/­Grounding_LLMs_with_online_RL
• Publication:
  hal-03970122‌​‌
• Contact:
  Clément Romac

7.1.17​​ SBMLtoODEjax

• Keywords:
  SBML, JAX,​​​‌ Python, Numerical simulations, Numerical‌ optimization, Automatic differentiation, Ordinary‌​‌ differential equations, Biomedical data​​
• Scientific Description:
  Advances in​​​‌ bioengineering and biomedicine demand‌ a deep understanding of‌​‌ the dynamic behavior of​​ biological systems, ranging from​​​‌ protein pathways to complex‌ cellular processes. Biological networks‌​‌ like gene regulatory networks​​ and protein pathways are​​​‌ key drivers of embryogenesis‌ and physiological processes. Comprehending‌​‌ their diverse behaviors is​​ essential for tackling diseases,​​​‌ including cancer, as well‌ as for engineering novel‌​‌ biological constructs. Despite the​​ availability of extensive mathematical​​​‌ models represented in Systems‌ Biology Markup Language (SBML),‌​‌ researchers face significant challenges​​ in exploring the full​​​‌ spectrum of behaviors and‌ optimizing interventions to efficiently‌​‌ shape those behaviors. Existing​​ tools designed for simulation​​​‌ of biological network models‌ are not tailored to‌​‌ facilitate interventions on network​​ dynamics nor to facilitate​​​‌ automated discovery. Leveraging recent‌ developments in machine learning‌​‌ (ML), this paper introduces​​ SBMLtoODEjax, a lightweight library​​​‌ designed to seamlessly integrate‌ SBML models with ML-supported‌​‌ pipelines, powered by JAX.​​ SBMLtoODEjax facilitates the reuse​​​‌ and customization of SBML-based‌ models, harnessing JAX's capabilities‌​‌ for efficient parallel simulations​​ and optimization, with the​​​‌ aim to accelerate research‌ in biological network analysis.‌​‌
• Functional Description:
  SBMLtoODEjax extends​​ SBMLtoODEpy, a python library​​​‌ developed in 2019 for‌ converting SBML files into‌​‌ python files written in​​ Numpy/Scipy. The chosen conventions​​​‌ for the generated variables‌ and modules are slightly‌​‌ different from the standard​​ SBML conventions (used in​​​‌ the SBMLtoODEpy library) with‌ the aim here to‌​‌ accommodate for more flexible​​ manipulations while preserving JAX-like​​​‌ functional programming style.
• URL:‌
  https://­developmentalsystems.­org/­sbmltoodejax/
• Publication:
  hal-04317246
• Contact:‌​‌
  Mayalen Etcheverry
• Partner:
  Tufts​​ University

7.1.18 Vivarium

• Name:​​​‌
  Large-scale simulator for research‌ and teaching in Artificial‌​‌ Intelligence and Artificial Life​​
• Keywords:
  Simulation, Artificial intelligence,​​​‌ Artificial Life, Multi-Agents System,‌ Teaching of programming, Research‌​‌
• Functional Description:

  This project​​ aims to seize these​​​‌ opportunities through the design‌ and implementation of a‌​‌ software platform providing an​​ integrated simulation environment for​​​‌ research, teaching, and dissemination‌ in the fields of‌​‌ Artificial Intelligence (AI) and​​ Artificial Life (AL). The​​​‌ project is titled The‌ Vivarium, which reflects a‌​‌ fundamental aspect of the​​​‌ convergence between these two​ domains: the emergence of​‌ complex behaviors, whether in​​ the natural or artificial​​​‌ world, necessarily relies on​ a need for adaptation​‌ to a complex environment​​ in which many autonomous​​​‌ entities interact.

  It will​ be used as an​‌ educational software in a​​ course from CISC Master​​​‌ at UPF-Barcelona in January​ 2025.

• Release Contributions:

  This​‌ release corresponds to the​​ state of the repo​​​‌ after all fixes were​ made following the SDIC​‌ course at Universitat Pompeu​​ Fabra of Barcelone (CSIM​​​‌ Master) in January 2025​ .

  This version mostly​‌ focuses on educational purposes,​​ with ready-to-use practical sessions​​​‌ in notebooks/sessions. Corentin Léger​ was the main contributor​‌ over the last year.​​

• News of the Year:​​​‌
  Corentin Léger, ingénieur de​ recherche recruté sur l'ANR​‌ JCJC ECOCURL (porté par​​ Clément Moulin-Frier) a mené​​​‌ un gros travail de​ développement du logiciel au​‌ cours de l'année 2024.​​ Ses applications pour l'enseignement​​​‌ sont maintenant validées. Le​ logiciel a notamment été​‌ utilisé pendant 10 heures​​ de travaux pratiques dans​​​‌ le Master CSIC de​ Universitat Pompeu Fabra à​‌ Barcelone, Espagne.
• URL:
  https://­github.­com/­flowersteam/­vivarium​​
• Contact:
  Clément Moulin-Frier
• Participant:​​​‌
  3 anonymous participants

7.1.19​ LLM_Culture

• Keywords:
  LLM, Multi-Agents​‌ System, Natural language processing​​
• Functional Description:

  Code for​​​‌ the 'Cultural evolution in​ populations of Large Language​‌ Models' paper. This repository​​ provides a comprehensive framework​​​‌ for studying the cultural​ evolution of linguistic content​‌ in populations of Large​​ Language Models (LLM).

  It​​​‌ allows organizing LLM agents​ into networks wherein each​‌ agent interacts with neighboring​​ agents by exchanging texts.​​​‌ Each agent can be​ assigned specific personalities and​‌ transmission instructions, serving as​​ prompts for generating new​​​‌ texts from their neighbors’​ narratives. Once the network​‌ structure and agent characteristics​​ are defined, you can​​​‌ simulate the cultural evolution​ of texts across generations​‌ of agents. We also​​ provide built-in metrics and​​​‌ vizualizations to analyze the​ results.

• URL:
  https://­github.­com/­flowersteam/­LLM-Culture
• Contact:​‌
  Jeremy Perez
• Participant:
  2​​ anonymous participants

7.1.20 TelephoneGameLLMs​​​‌

• Keywords:
  Large Language Models,​ Multi-Agents System, Cultural Evolution​‌
• Functional Description:
  Code for​​ the paper "When LLMs​​​‌ Play the Telephone Game:​ Cumulative Changes and Attractors​‌ in Iterated Cultural Transmissions"​​ https://arxiv.org/abs/2407.04503 In this paper,​​​‌ we introduce conceptual and​ methodological tools for evaluating​‌ Large Language Models in​​ multi-turn settings. Those tools​​​‌ are inspired by cultural​ evolutionary theory, and in​‌ particular by the concepts​​ of cultural attractors.
• URL:​​​‌
  https://­github.­com/­flowersteam/­TelephoneGameLLM
• Publication:
  hal-04714994
• Contact:​
  Jeremy Perez

7.1.21 styr​‌

• Name:
  Stick To Your​​ Role
• Keywords:
  LLM, Cognitive​​​‌ sciences
• Functional Description:

  Code​ for our paper https://arxiv.org/abs/2402.14846​‌ and leaderboard https://huggingface.co/spaces/flowers-team/StickToYourRoleLeaderboard.

  Enables​​ evaluating LLMs using personal​​​‌ value questionnaires (PVQ, SVS).​ More precisely, it instructs​‌ the LLM to simulated​​ various personas and exposes​​​‌ it to different contexts​ (e.g. long reddit posts).​‌ Then it evaluates the​​ value stability of the​​​‌ simulated population between those​ contexts. Additionally, it computes​‌ confirmatory factor analysis (CFI,​​ SRMR, RMSEA), and the​​​‌ structure of expressed values​ (stress metric).

• URL:
  https://­github.­com/­flowersteam/­value_stability​‌
• Contact:
  Grgur Kovac

7.1.22​​ transformerXL_PPO_JAX

• Keywords:
  Reinforcement learning,​​​‌ Transformer
• Functional Description:

  This​ repository provides a JAX​‌ implementation of TranformerXL with​​ PPO in a RL​​ setup following : "Stabilizing​​​‌ Transformers for Reinforcement Learning"‌ from Parisotto et al.‌​‌ (https://arxiv.org/abs/1910.06764).

  The code uses​​ the PureJaxRL template for​​​‌ PPO and copied some‌ of the code from‌​‌ hugging face trasnformer XL​​ repo transferring it to​​​‌ JAX. We also took‌ inspiration from the pytorch‌​‌ code in https://github.com/MarcoMeter/episodic-transformer-memory-ppo, which​​ has some simplification of​​​‌ gradient propagation and positional‌ encoding compared to transformerXL‌​‌ as it is described​​ in the original paper​​​‌ (https://arxiv.org/abs/1901.02860). The training handles‌ [Gymnax](https://github.com/RobertTLange/gymnax) environment.

  We also‌​‌ tested it on Craftax​​, on which it​​​‌ beat the baseline presented‌ in the paper (https://arxiv.org/abs/2402.16801)‌​‌ including PPO-RNN, training with​​ unsupervised environment design and​​​‌ intrinsic motivation. Notably we‌ reach the 3rd level‌​‌ (the sewer) and obtain​​ several advanced advancements, which​​​‌ was not achieved by‌ the methods presented in‌​‌ the paper. See Craftax​​ Results for more informations.​​​‌

  The training of a‌ 5M transformer on craftax‌​‌ for 1e9 steps (with​​ 1024 environments) takes about​​​‌ 6h30 on a single‌ A100.

• Contact:
  Gautier Hamon‌​‌

7.1.23 ER-MRL

• Keywords:
  Reinforcement​​ learning, Evolutionary Algorithms, Recurrent​​​‌ network
• Functional Description:

  Code‌ for the "Evolving-Reservoirs-for-Meta-Reinforcement-Learning" (ER-MRL)‌​‌ paper (https://arxiv.org/abs/2312.06695).

  We adopt​​ a computational framework based​​​‌ on meta reinforcement learning,‌ modeling the interplay between‌​‌ evolution and development. At​​ the evolutionary scale, we​​​‌ evolve reservoirs, a family‌ of recurrent neural networks‌​‌ generated from hyperparameters. These​​ evolved reservoirs are then​​​‌ utilized to facilitate the‌ learning of a behavioral‌​‌ policy through reinforcement learning.​​ This is done by​​​‌ encoding the environment state‌ through the reservoir before‌​‌ providing it to the​​ agent's policy.

• Contact:
  Corentin​​​‌ Leger

7.1.24 LLM4Humanities

• Keywords:‌
  LLM, Python, Data Generator,‌​‌ Generative AI
• Scientific Description:​​
  Qualitative research in experimental​​​‌ psychology and the humanities‌ often relies on manual‌​‌ annotation of textual data​​ using defined codebooks. This​​​‌ process is indispensable but‌ time-consuming and costly. Moreover,‌​‌ best practices require at​​ least two independent annotators​​​‌ in order to compute‌ inter-rater reliability (IRR), which‌​‌ further increases the required​​ resources. IRR is crucial​​​‌ to distinguish variance due‌ to coder subjectivity from‌​‌ variance due to the​​ phenomenon under study, yet​​​‌ in practice it is‌ frequently omitted, misreported, or‌​‌ computed using inadequate metrics​​ (e.g., raw percentage agreement​​​‌ or simple correlations). The‌ objective of the LLM4Humanities‌​‌ project is to design​​ an open-source, Python-based toolkit​​​‌ and web application that‌ leverages large language models‌​‌ (LLMs) to support, accelerate,​​ and improve the methodological​​​‌ rigor of qualitative annotation‌ workflows. In addition to‌​‌ annotation assistance, LLM4Humanities includes​​ a generation mode designed​​​‌ to support the creation‌ of experimental material. In‌​‌ this mode, users can​​ select one or several​​​‌ template items (e.g., a‌ mathematics exercise) and specify‌​‌ a set of constraints.​​ The system then generates​​​‌ multiple new variants of‌ the item. These generated‌​‌ items can subsequently be​​ passed through the same​​​‌ annotation and evaluation pipeline,‌ providing a first automated‌​‌ assessment of the quality​​ and consistency of the​​​‌ generated content.
• Functional Description:‌
  LLM4Humanities is an open-source,‌​‌ Python-based toolkit and web​​ application that integrates LLM-assisted​​​‌ annotation, inter-rater reliability analysis,‌ and the generation of‌​‌ controlled variants of experimental​​​‌ material within a single​ end-to-end workflow
• URL:
  https://­github.­com/­flowersteam/­LLM4Humanities​‌
• Contact:
  Olivier Clerc
• Participant:​​
  Olivier Clerc

7.2.1 ToGather application​

Participants: Cécile Mazon,​‌ Hélène Sauzéon, Eric​​ Meyer, Isabeau Saint-Supery​​​‌.

• Name:
  Application for​ Specialized education
• Keywords:
  Parent-professional​‌ relationships; user-centered design; school​​ inclusion; autism spectrum disorder;​​​‌ ecosystemic approach
• Participants:
  Isabeau​ Saint-supery, Cécile Mazon, Hélène​‌ Sauzéon, Agilonaute
• Scientific Description:​​
  With participatory design methods,​​​‌ we have designed an​ interactive website application for​‌ educational purposes. This application​​ aims to provide interactive​​​‌ services with continuously updated​ content for the stakeholders​‌ of school inclusion of​​ children with specific educational​​​‌ needs. Especially, the services​ provide: 1) the student's​‌ profile with strengths and​​ weaknesses; 2) an evaluation​​​‌ and monitoring over time​ of the student's repertoire​‌ of acquired, emerging or​​ targeted skills; 3) a​​​‌ shared notebook of effective​ psycho-educational solutions for the​‌ student ; 4) a​​ shared messaging system for​​​‌ exchanging "news" about the​ student and his/her family​‌ and, 5) a meeting​​ manager allowing updates of​​​‌ evaluations (student progress). This​ application is currently assessed​‌ with a field study.​​ Then, it will be​​​‌ transferred to the Academy​ of Nouvelle-Aquitaine-Bordeaux of the​‌ National Education Ministery.
• URL:​​
  The website is not​​​‌ online yet, but all​ informations such as tutorials​‌ are here.
• Publication:​​
  hal-03436355

8.1.1 ACES:​‌ Generating a Diversity of​​ Challenging Programming Puzzles with​​​‌ Autotelic Generative Models

Participants:​ Julien Pourcel [correspondant],​‌ Cédric Colas, Gaia​​ Molinaro, Pierre-Yves Oudeyer​​​‌, Laetitia Teodorescu.​

Results. 

We report‌​‌ results of our runs​​ in Figure 3.​​​‌ Overall, the methods based‌ on semantic archives, ACES‌​‌ and ELM-Semantic, achieve the​​ highest diversity in the​​​‌ semantic space. We report‌ diversity measures inside the‌​‌ embedding spaces of various​​ smaller language models in​​​‌ Figure 4. In‌ these figures we see‌​‌ that overall ACES outperforms​​ other methods in this​​​‌ measure of diversity. We‌ additionally perform tests of‌​‌ the suitability of generated​​ puzzles as finetuning data​​​‌ for smaller LMs. For‌ all methods, we finetune‌​‌ a smaller model (​​OpenLlama-3b) on the​​​‌ generated set and we‌ test the pass@k metric‌​‌ for different values of​​ k on the P3​​​‌ test set; we report‌ the scores in Figure‌​‌ 5. From that​​ figure we see that​​​‌ we encounter a tradeoff‌ between how diverse the‌​‌ data is and how​​​‌ useful it is to​ get a high score​‌ on the P3 test​​ set. Further work is​​​‌ needed to get data​ that is both diverse​‌ and useful.a

Overview of​​ ACES. ACES maintains an​​​‌ archive of discovered puzzles​ grouped into cells indexed​‌ by their semantic representation​​ (skill combination). ACES runs​​​‌ in several steps: 1)​ sample a target semantic​‌ goal and relevant examples​​ from the archive. 2)​​​‌ given these, generate a​ puzzle f and its​‌ solution g with the​​ puzzle generator. 3) test​​​‌ the validity of that​ pair by running assert(f(g())​‌ in the interpreter. 4)​​ if the pair is​​​‌ valid, obtain its semantic​ representation with the puzzle​‌ labeler. 5) add the​​ new pair to its​​​‌ corresponding cell in the​ archive.

Diversity​ of generated puzzles in​‌ semantic space. We report​​ the evolution of several​​​‌ diversity metrics computed in​ the semantic space as​‌ a function of the​​ number of puzzle-solution pairs​​​‌ generated by the puzzle​ generator. Semantic algorithms (algname​‌ and ELM semantic) achieve​​ higher diversity in the​​​‌ semantic space.

Diversity of generated puzzles​‌ in embedding spaces. We​​ report the evolution of​​​‌ the pairwise distance between​ puzzle-solution pair embeddings as​‌ a function of the​​ number of generated puzzle-solution​​​‌ pairs, for three different​ embedding representation spaces (average​‌ across seeds).

Downstream​ performance on the P3​‌ test set. Pass@k is​​ the fraction of puzzles​​​‌ solved after $k$ attempts​ ($k\in$[1:10]).​‌ Green overlaps with yellow.​​

8.1.2 MAGELLAN: Metacognitive Generalization​​ of Learning Progress for​​​‌ Online RL in LLM​ agents

Participants: Loris Gaven​‌ [correspondant], Thomas Carta​​, Clément Romac,​​​‌ Cédric Colas, Pierre-Yves​ Oudeyer, Olivier Sigaud​‌ [ISIR Sorbonne Université, Paris,​​ France], Sylvain Lamprier​​ [Univ Angers, LERIA].​​​‌

We are developing MAGELLAN‌47, a method‌​‌ designed to enable LLM-based​​ reinforcement learning (RL) agents​​​‌ to estimate their own‌ Learning Progress (LP) and‌​‌ use it to dynamically​​ organize their training curriculum.​​​‌ By leveraging the LLM's‌ rich semantic representations, MAGELLAN‌​‌ allows agents to generalize​​ LP estimations to unseen,​​​‌ language-defined goals, overcoming limitations‌ of classical methods that‌​‌ require direct evaluation of​​ each goal.

MAGELLAN uses​​​‌ the LLM to generate‌ latent representations of goals‌​‌ and tasks, capturing their​​ semantic relationships. It continuously​​​‌ monitors the agent's performance‌ over time, estimating LP‌​‌ as the change in​​ success rates for specific​​​‌ goals. This approach enables‌ the agent to identify‌​‌ goals where it is​​ making progress and focus​​​‌ its training on those‌ areas. MAGELLAN's integration ensures‌​‌ that the LLM-based agent​​ can simultaneously refine its​​​‌ policy and competence estimations,‌ adapting both to new‌​‌ tasks in real time.​​

Our experiments in the​​​‌ Little-Zoo environment, which‌ features hierarchical and commonsense-driven‌​‌ tasks, demonstrate that MAGELLAN​​ effectively prioritizes high-LP goals,​​​‌ even when faced with‌ novel or unseen tasks.‌​‌ Unlike traditional LP estimation​​ methods, which rely on​​​‌ direct evaluations and struggle‌ with generalization, MAGELLAN enables‌​‌ the agent to quickly​​ identify meaningful learning opportunities.​​​‌ This results in faster‌ adaptation, improved sample efficiency,‌​‌ and more effective curriculum​​ organization, paving the way​​​‌ for truly autonomous agents‌ capable of navigating vast‌​‌ and complex goal spaces.​​

8.1.3 When goals are​​​‌ beyond reach: Metacognitive monitoring‌ guides autonomous discovery of‌​‌ frugal assistance-seeking in LLMs​​

Participants: Clément Romac [correspondant]​​​‌, Pierre-Yves Oudeyer.‌

Enhancing LLMs with metacognitive‌​‌ capabilities has been identified​​ as a key challenge​​​‌ for improving the trustworthiness‌ and interpretability of these‌​‌ models. In this work,​​ we investigate how such​​​‌ metacognitive abilities can be‌ leveraged to trigger external‌​‌ assistance when the model’s​​ own capabilities are insufficient.​​​‌ While improving LLMs’ is‌ essential, it is equally‌​‌ critical that models learn​​ to recognize their own​​​‌ limitations—and to seek or‌ rely on external support‌​‌ in real-world settings where​​ functional competence may be​​​‌ partial or underdeveloped. This‌ ability forms a crucial‌​‌ part of a broader​​ learning loop: requesting help​​​‌ when needed, then internalizing‌ the knowledge or skills‌​‌ acquired through that assistance.​​

Augmenting LLMs with external​​​‌ assistance and, in particular,‌ what has been named‌​‌ "tools", has become a​​ well-established practice. These augmentations​​​‌ range from calculators and‌ retrieval systems to code‌​‌ interpreters, and even other​​ LLMs. This shift has​​​‌ led to a rethinking‌ of the role of‌​‌ LLMs—not as general-purpose solvers,​​ but as assistants (often​​​‌ referred to as action‌ models) that must learn‌​‌ to orchestrate the use​​ of external resources and​​​‌ integrate their outputs into‌ coherent, human-readable responses.

This‌​‌ reframing introduces a new​​ class of decision-making problems:​​​‌ LLMs must determine when‌ and which external assistance‌​‌ to invoke. However, the​​ optimal assistance strategy is​​​‌ not known in advance.‌ Some tasks may be‌​‌ solvable independently by the​​ LLM, while others may​​​‌ require external help. Additionally,‌ the tools themselves may‌​‌ be fallible—for instance, even​​​‌ large or specialized LLMs​ can return suboptimal results.​‌ To address this, most​​ prior approaches rely on​​​‌ supervised learning, fine-tuning LLMs​ on curated datasets containing​‌ examples of effective tool​​ use. More recently, several​​​‌ works have begun exploring​ how RL can be​‌ used to learn assistance-seeking​​ strategies from scratch, without​​​‌ requiring predefined tool-use demonstrations.​

While both RL and​‌ more conventional supervised learning​​ approaches have shown promise,​​​‌ an important dimension of​ the assistance-seeking problem remains​‌ largely understudied: external assistance​​ comes at a cost.​​​‌ This cost may take​ the form of increased​‌ latency in the LLM’s​​ response, financial charges for​​​‌ calling APIs, or computational​ overhead. Although early work​‌ on tool use—such as—acknowledged​​ this issue, it has​​​‌ received limited attention since.​ A recent exception is,​‌ which introduced a first​​ approach to this multi-objective​​​‌ problem: maximizing task performance​ while minimizing assistance costs.​‌ Their method involves a​​ multi-stage learning pipeline: (1)​​​‌ an estimator of LLM​ performance is trained using​‌ interaction data between the​​ LLM and the task​​​‌ space; (2) a separate​ model is trained to​‌ simulate the outputs of​​ both the LLM and​​​‌ its assistance sources; and​ (3) given a predefined​‌ cost budget, Dynamic Programming​​ is used to derive​​​‌ the optimal assistance strategy.​ While effective, this method​‌ is computationally intensive and​​ requires extensive data collection​​​‌ and training across multiple​ stages.

In this work,​‌ we propose a fully​​ online approach (see Figure​​​‌ 6) based on​ multi-objective contextual multi-armed bandits.​‌ Given a task, we​​ frame the decision of​​​‌ whether to keep the​ task with the LLM​‌ or delegate it to​​ external assistance as the​​​‌ selection of an arm.​ Given a task, we​‌ consider the dual objective​​ of maximizing the answer's​​​‌ performance $R$ while minimizing​ its cost $C$,​‌ and we adopt scalarization—i.e.,​​ combining the two objectives​​​‌ into a single weighted​ sum $U=\mathrm{\beta ‌}R+(\mathrm{1}-\beta )\mathrm{C‌}$ that our approach aims​ to maximize. Crucially, our​‌ method naturally adapts to​​ any specified user-specified budget​​​‌ by treating the budget​ as the scalarization weight​‌ that balances the two​​ objectives. A central challenge​​​‌ of this approach lies​ in efficiently estimating the​‌ performance and cost associated​​ with each option (i.e.,​​​‌ the LLM and all​ available assistance sources), using​‌ as few interactions as​​ possible. To address this,​​​‌ we draw inspiration from​ MAGELLAN and leverage the​‌ LLM itself to learn​​ these estimations.

We first​​​‌ evaluate our method on​ a set of carefully​‌ designed math problems with​​ calculator tools as assistance,​​​‌ for which the optimal​ strategy is known. This​‌ setting enables us to​​ investigate how the strategy​​​‌ discovered by our method​ compares to the optimal​‌ one, as well as​​ the sample efficiency of​​​‌ our approach (i.e., the​ number of interactions required​‌ to converge). We notably​​ show that our LLM-based​​​‌ estimation of performance and​ cost reaches similar or​‌ even better performance than​​ a classic moving average​​​‌ approach which has access​ to privileged information—namely, the​‌ problem category. Finally, we​​ demonstrate the broader applicability​​ of our method by​​​‌ applying it to real-world‌ problems faced by LLMs.‌​‌ In particular, we apply​​ it to a standard​​​‌ question-answering benchmark: MMLU-Pro. The‌ results show that our‌​‌ approach is scalable to​​ complex natural language tasks​​​‌ without access to any‌ external expert knowledge.

We‌​‌ study how LLMs can​​ autonomously learn how and​​​‌ when to use tools‌ when solving tasks.

8.1.4 LLM-based goal generation​​​‌ for autotellic exploration with‌ goal-conditioned RL

Participants: Guillaume‌​‌ Pourcel [correspondant], Grgur​​ Kovač, Thomas Carta​​​‌, Cédric Colas,‌ Pierre-Yves Oudeyer.

Designing‌​‌ autotelic agents capable of​​ autonomously generating and pursuing​​​‌ their own goals represents‌ a promising endeavor for‌​‌ open-ended learning and skill​​ acquisition in reinforcement learning.​​​‌ This challenge is especially‌ difficult in open worlds‌​‌ that require inventing new​​ previously unobserved goals. In​​​‌ this work, we propose‌ an architecture where a‌​‌ single generalist autotelic agent​​ is trained on an​​​‌ automatic curriculum of goals.‌ We leverage large language‌​‌ models (LLMs) to generate​​ goals as code for​​​‌ reward functions based on‌ learnability and difficulty estimates.‌​‌ The goal-conditioned RL agent​​ is trained on those​​​‌ goals sampled based on‌ learning progress. We compare‌​‌ our method to an​​ adaptation of OMNI-EPIC to​​​‌ goal-conditioned RL. Our preliminary‌ experiments imply that our‌​‌ method generates a higher​​ proportion of learnable goals,​​​‌ suggesting better adaptation to‌ the goalconditioned learner. This‌​‌ work is described in​​ this technical report.​​​‌

The used architecture.

8.1.5 Self-Improving Language​​ Models for Evolutionary Program​​​‌ Synthesis: A Case Study‌ on ARC-AGI

Participants: Julien‌​‌ Pourcel [correspondant], Cédric​​ Colas, Pierre-Yves Oudeyer​​​‌.

In our work‌ on SOAR (Self-improving Operators‌​‌ for Automated program Refinements),​​ published at ICML 2025,​​​‌ we address a fundamental‌ limitation in program synthesis:‌​‌ while large language models​​ struggle to solve complex​​​‌ tasks in single attempts,‌ traditional evolutionary approaches are‌​‌ constrained by the fixed​​ capabilities of their underlying​​​‌ generative models. We developed‌ a framework that integrates‌​‌ language models into a​​ self-improving evolutionary loop, enabling​​​‌ continuous performance enhancement through‌ experience rather than relying‌​‌ on static model capabilities.​​

SOAR architecture.

Our‌ method operates through an‌​‌ iterative two-phase process that​​ we designed to create​​​‌ a virtuous cycle of‌ improvement. First, an evolutionary‌​‌ search phase employs a​​ language model to sample​​​‌ and refine candidate program‌ solutions. Second, a hindsight‌​‌ learning phase converts these​​ search attempts—both successful and​​​‌ unsuccessful—into valid problem-solution pairs‌ that we use to‌​‌ fine-tune the LLM's sampling​​ and refinement capabilities. This​​​‌ approach leverages positive transfer‌ between the sampling and‌​‌ refinement fine-tuning tasks, allowing​​​‌ the system to bootstrap​ its own improvement without​‌ requiring human-engineered training data.​​

We evaluated SOAR on​​​‌ the challenging ARC-AGI benchmark,​ which tests abstract reasoning​‌ and program induction capabilities.​​ Our framework solves 52%​​​‌ of the public test​ set, establishing state-of-the-art results​‌ for program synthesis using​​ open-source language models. These​​​‌ improvements compound through iterations,​ with models showing enhanced​‌ abilities to both generate​​ initial program ideas and​​​‌ refine existing solutions. Notably,​ the gains carry over​‌ to test-time adaptation, enabling​​ continuous improvement on target​​​‌ problems even after deployment.​

Our research demonstrates that​‌ program synthesis systems can​​ transcend the limitations of​​​‌ their base models through​ self-improvement, opening new possibilities​‌ for autonomous AI development.​​ By showing how iterative​​​‌ model improvement can overcome​ performance plateaus inherent to​‌ search methods, SOAR provides​​ a drop-in upgrade for​​​‌ existing systems like AlphaEvolve​ or ShinkaEvolve, transforming their​‌ fixed LLM operators into​​ continuously improving ones.

8.1.6​​​‌ WorldLLM: Improving LLMs' world​ modeling using curiosity-driven theory-making​‌

Participants: Guillaume Levy,​​ Cedric Colas, Pierre-Yves​​​‌ Oudeyer, Thomas Carta​, Cément Romac [correspondant]​‌.

Large Language Models​​ (LLMs) possess broad knowledge​​​‌ about the world, but​ leveraging this knowledge for​‌ precise dynamics modeling remains​​ challenging. While LLMs can​​​‌ engage in general reasoning,​ they struggle to make​‌ accurate predictions in specific​​ domains with structured observations​​​‌ and dynamics, such as​ physics simulations or video​‌ games. This limitation stems​​ from the gap between​​​‌ their general capabilities and​ the need for grounded,​‌ domain-specific understanding.

In this​​ paper, we present WorldLLM​​​‌59, a framework​ for autonomous improvement of​‌ an LLM's world modeling​​ abilities. Our approach combines​​​‌ 1) probabilistic theory induction​ to produce hypotheses that​‌ are given in our​​ LLM's prompt to improve​​​‌ its predictions and 2)​ curiosity-driven RL to explore​‌ the environment and collect​​ transitions poorly predicted by​​​‌ the current hypotheses (see​ Figure 9). Formally,​‌ our LLM's world model​​ is the conditional probability​​​‌ $P\left(s_{\mathrm{t}+1}\right|{\mathrm{s‌}}_t,a_{\mathrm{t}},H)$,​​​‌ where $s_t$ represents​ a state, $a_{\mathrm{t‌}}$ an action, and $\mathrm{H}$ a set of natural​​​‌ language hypothesized theories. This​ probability is computed by​‌ the LLM by giving​​ it $s_t$,​​​‌ $a_t$, and​ $H$ in its prompt​‌ and taking the probability​​ of $s_{t+‌1}$ to follow this​ prompt. Our key insight​‌ is that natural language​​ theories can help ground​​​‌ an LLM's broad knowledge​ into precise predictive power​‌ by providing domain-specific rules.​​ Our approach consists of​​​‌ three interacting components: (1)​ our LLM that computes​‌ $P\left(s_{\mathrm{t}+1}\right|{\mathrm{s‌}}_t,a_{\mathrm{t}},H)$ by​‌ conditioning its predictions on​​ both a state-action pair​​​‌ and the current hypotheses,​ (2) a theory generator​‌ that updates natural language​​ hypotheses using Bayesian inference,​​​‌ and (3) a curiosity-driven​ reinforcement learning agent trained​‌ to collect evidence against​​ the current hypotheses. Inspired​​​‌ by how humans, from​ children to scientists, actively​‌ update their internal world​​ model by performing experiments,​​ our agent's exploration provides​​​‌ new evidence for hypothesis‌ refinement, creating a virtuous‌​‌ cycle of improvement.

WorldLLM.​​

We demonstrate our approach‌ in a video game‌​‌ environment where agents manipulate​​ and combine objects, showing​​​‌ that WorldLLM successfully learns‌ accurate predictive models while‌​‌ generating human-interpretable theories about​​ environment dynamics. This work​​​‌ contributes to a growing‌ body of research on‌​‌ improving LLMs' world modeling​​ capabilities and grounding their​​​‌ knowledge in specific domains.‌ By combining ideas from‌​‌ theory-based RL, Bayesian inference,​​ and active exploration, we​​​‌ provide a framework for‌ learning structured, interpretable world‌​‌ models that leverage both​​ the broad knowledge of​​​‌ LLMs and domain-specific experiences‌ without any costly gradient-based‌​‌ learning.

8.1.7 HERAKLES: Hierarchical​​ Skill Compilation for Open-ended​​​‌ LLM Agents

Participants: Thomas‌ Carta [correspondant], Cément‌​‌ Romac, Loris Gaven​​, Pierre-Yves Oudeyer,​​​‌ Olivier Sigaud, Sylvain‌ Lamprier.

In our‌​‌ work on HERAKLES (HiERarchicAl​​ sKill compiLation for open-Ended​​​‌ llm agentS), we address‌ a fundamental challenge in‌​‌ open-ended AI: as goal​​ spaces expand, increasingly complex​​​‌ goals require composing multiple‌ elementary actions, leading to‌​‌ combinatorial explosion that impedes​​ learning progress. While existing​​​‌ hierarchical reinforcement learning approaches‌ rely on expert-defined skill‌​‌ spaces and pre-trained low-level​​ policies, such designs are​​​‌ inadequate for open-ended scenarios‌ where goal spaces naturally‌​‌ diversify across a broad​​ spectrum of difficulties. We​​​‌ developed a framework that‌ enables continuous skill compilation,‌​‌ dynamically expanding the agent's​​ capabilities through experience rather​​​‌ than relying on fixed,‌ predefined abstractions.

Our method‌​‌ operates through a two-level​​ hierarchical architecture designed to​​​‌ create a virtuous cycle‌ of skill acquisition. A‌​‌ high-level policy, instantiated as​​ a Large Language Model,​​​‌ decomposes complex goals into‌ subgoals and selects skills‌​‌ from an evolving skill​​ space. A low-level policy,​​​‌ implemented as lightweight neural‌ networks, executes these skills‌​‌ through primitive actions. Crucially,​​ as the hierarchical agent​​​‌ masters a goal, the‌ complete trajectory is compiled‌​‌ into the low-level policy​​ as a new reusable​​​‌ skill. A competence estimator‌ predicts the low-level policy's‌​‌ success probability for each​​ skill, ensuring the high-level​​​‌ policy only invokes skills‌ that can be reliably‌​‌ executed. This approach leverages​​ language's compositional and combinatorial​​​‌ properties to structure the‌ skill space, enabling generalization‌​‌ across semantically related goals.​​

We evaluated HERAKLES in​​​‌ the Crafter environment, a‌ procedurally generated Minecraft-like world‌​‌ designed to assess agent​​ capabilities within a unified​​​‌ open-ended framework. Our framework‌ achieves Crafter scores above‌​‌ 70, while baselines plateau​​ below 30. More importantly,​​​‌ HERAKLES scales near-linearly with‌ goal difficulty, whereas non-hierarchical‌​‌ methods exhibit exponential growth​​ in learning time. The​​​‌ framework also demonstrates strong‌ generalization: when tested on‌​‌ synonymous goal formulations, HERAKLES​​ shows only a 16%​​​‌ performance drop compared to‌ 24-27% for baselines, and‌​‌ maintains robust performance on​​ compositional variants requiring repeated​​​‌ skill execution.

Our research‌ demonstrates that open-ended agents‌​‌ can transcend the limitations​​ of fixed skill spaces​​​‌ through continuous compilation, opening‌ new possibilities for lifelong‌​‌ learning systems. By showing​​ how mastered behaviors can​​​‌ be recursively encoded at‌ lower levels for rapid‌​‌ reuse—mirroring how humans overcome​​​‌ complexity barriers through hierarchical​ learning—HERAKLES provides a principled​‌ approach for building agents​​ that autonomously expand their​​​‌ competencies over time.

8.1.8​ Software Engineering Agents for​‌ Embodied Controller Generation :​​ A Study in Minigrid​​​‌ Environments

Participants: Timothé Boulet​ [correspondant], Xavier Hinaut​‌ [Mnemosyne, Inria Bordeaux],​​ Clément Moulin-Frier, Nathanaël​​​‌ Fijalkow.

Motivation. 

Software​ Engineering Agents (SWE-Agents) have​‌ proven effective for traditional​​ software engineering tasks with​​​‌ accessible codebases, but their​ performance for embodied tasks​‌ requiring well-designed information discovery​​ remains unexplored. In this​​​‌ paper 44, we​ present the first extended​‌ evaluation of SWE-Agents on​​ controller generation for embodied​​​‌ tasks, adapting Mini-SWE-Agent (MSWEA)​ to solve 20 diverse​‌ embodied tasks from the​​ Minigrid environment. Our experiments​​​‌ compare agent performance across​ different information access conditions:​‌ with and without environment​​ source code access, and​​​‌ with varying capabilities for​ interactive exploration. We quantify​‌ how different information access​​ levels affect SWE-Agent performance​​​‌ for embodied tasks and​ analyze the relative importance​‌ of static code analysis​​ versus dynamic exploration for​​​‌ task solving. This work​ establishes controller generation for​‌ embodied tasks as a​​ crucial evaluation domain for​​​‌ SWE-Agents and provides baseline​ results for future research​‌ in efficient reasoning systems.​​

This work investigates a​​​‌ fundamental question: How do​ SWE-Agents perform in controller​‌ generation for embodied tasks​​ ? Our approach involves​​​‌ a code-agent (the SWE-Agent​ interacting with a code-environment​‌ involving codebases and terminals)​​ that generates controller-agents (Python​​​‌ programs) to solve tasks​ in an embodied setup,​‌ creating a two-level agency​​ structure that differs from​​​‌ direct LLM-environment interaction approaches​ common in embodied AI.​‌ Figure 10 illustrates this​​ two-level agency structure. The​​​‌ agent can evaluate its​ proposed solution by executing​‌ them in the environment​​ and receiving feedback in​​​‌ the form of success/failure​ and reward. Task terminates​‌ either when the agent​​ validates with a special​​​‌ command or when the​ maximum number of steps​‌ or cost is reached.​​

Two-level agency structure: a​​​‌ code-agent interacts with a​ code-environment to generates controller-agents​‌ (Python programs) to solve​​ the embodied task.

We evaluate the​‌ challenge of controller generation​​ for embodied tasks by​​​‌ adapting Mini-SWE-Agent (MSWEA) to​ solve diverse Minigrid tasks​‌ under different information access​​ conditions:

• Source Code Access​​​‌: When the agent​ can read Minigrid environment​‌ code, it can analyze​​ environment mechanics, constraints, and​​​‌ object interactions to inform​ controller design.
• Interactive Exploration​‌: When the agent​​ can write and execute​​​‌ scripts to probe the​ environment, where it can​‌ discover dynamics through exploration,​​ observing outcomes of actions​​​‌ in various states.

Results.​ 

The best@5 success rates​‌ of MSWEA across different​​ tasks and information access​​​‌ conditions are summarized in​ Figure 11. We​‌ display standard deviation as​​ error bars in all​​​‌ our plots.

Minigrid PO​ was very hard to​‌ solve for MSWEA, with​​ many tasks not being​​​‌ solved even with full​ access. In Minigrid FO​‌ however, all tasks except​​ 1 are solved by​​ at least MSWEA with​​​‌ full access. Partial Observability,‌ as a component of‌​‌ embodied tasks, is thus​​ a hard step for​​​‌ SWE Agents to solve.‌

Best@5 success rate of‌​‌ MSWEA across different tasks​​ and information access conditions​​​‌ in Fully Observable Minigrid.‌

To‌​‌ identify patterns in the​​ influence of the type​​​‌ of the task to‌ the performance of different‌​‌ information access conditions, we​​ grouped the average best@5​​​‌ success rate metric into‌ 4 categories : navigation,‌​‌ manipulation, hazard, memory, as​​ well as the overall​​​‌ average across all tasks.‌ The results are shown‌​‌ in Figures 12 and​​ Figure 13.

Mean-by-category​​​‌ best@5 success rate in‌ Fully Observable Minigrid

Mean-by-category best@5‌ success rate in Partially‌​‌ Observable Minigrid

In the Fully‌​‌ Observable benchmark, comparing MSWEA​​ (blue bars) with its​​​‌ fully ablated version (red‌ bars) without neither source‌​‌ code read access nor​​ interactive exploration, we observe​​​‌ performance drop dramatically. An‌ agent with only the‌​‌ Test-Access capability (i.e. being​​ capable of testing its​​​‌ solution to obtain the‌ success rate of its‌​‌ controller solution on the​​ task) obtain much worse​​​‌ result, but surprisingly still‌ manages to solve some‌​‌ tasks through iterated submissions.​​

If we try to​​​‌ get back to the‌ MSWEA performance level by‌​‌ adding only the code​​ access (cyan bars), we​​​‌ see very limited improvement,‌ which means reading only‌​‌ help partially the agent​​ and that the difficulty​​​‌ lies elsewhere. If we‌ add only the interactive‌​‌ execution capability however (orange​​ bars), we observe the​​​‌ performance get back to‌ a comparable level as‌​‌ MSWEA. This pattern is​​ consistent across all task​​​‌ categories and particularly for‌ manipulation task, where the‌​‌ very exact knowledge of​​ how the environment operates​​​‌ is required to solve‌ the task. This systematic‌​‌ pattern means that the​​ interactive access is an​​​‌ essential capacity of SWE-Agents‌ that allows them to‌​‌ perform significantly better in​​ embodied tasks.

In the​​​‌ Partially Observable benchmark, performance‌ is much lower than‌​‌ in Minigrid FO, in​​ particular for the complex​​​‌ manipulation tasks. We can‌ note there are different‌​‌ patterns depending on the​​ task category, but we​​​‌ will not try to‌ interpret them as these‌​‌ may arise either from​​ statistical variability given the​​​‌ relatively high standard errors,‌ or from subtle hard‌​‌ to infer and task-specific​​ factors that bias the​​​‌ agent’s behavior in ways‌ not observed in similar‌​‌ tasks. The overall performance​​ does not vary significantly​​​‌ with the information access‌ conditions. We interpret this‌​‌ as the PO tasks​​ being inherently too hard​​​‌ for MSWEA, such that‌ the agent only solve‌​‌ the simplest tasks such​​ as the easiest navigation​​​‌ tasks, and can make‌ little use of different‌​‌ information access to increase​​ performances. This leads us​​​‌ to believe that strongly‌ embodied tasks such as‌​‌ Minigrid PO tasks represent​​​‌ a good benchmark for​ SWE Agents : they​‌ perform decently on some​​ tasks, but on others,​​​‌ even with good LLMs​ and access to source​‌ code and execution access,​​ they still have significant​​​‌ room for improvement regarding​ the understanding of the​‌ functioning of the environment.​​ These results encourages the​​​‌ use of embodied tasks​ for future software engineering​‌ agents benchmarks.

8.2.1 The effect​ of social network structure​‌ on collective innovation

Participants:​​ Eleni Nisioti [correspondant],​​​‌ Mateo Mahaut, Pierre-Yves​ Oudeyer, Ida Momennejad​‌, Sebastian Risi,​​ Pierre-Yves Oudeyer, Clément​​​‌ Moulin-Frier.

Innovations are​ a central component of​‌ open-ended skill acquisition: they​​ denote the emergence of​​​‌ new solutions by the​ recombination of existing ones​‌ and their presence is​​ necessary to ensure a​​​‌ continuous complexification of an​ agent's cultural repertoire. While​‌ we often tend to​​ attribute discoveries to certain​​​‌ innovative individuals, if we​ shed a broad perspective​‌ at the history of​​ our species we see​​​‌ that human innovation is​ primarily a collective process.​‌ Fields such as psychology​​ and anthropology have been​​​‌ studying the ability of​ human groups to innovate​‌ for some time, with​​ studies indicating that the​​​‌ social network structure has​ a significant impact: fully-connected​‌ structures are better suited​​ for quick convergence in​​​‌ easy problems with clear​ global optima, while partially-connected​‌ structures perform best in​​ difficult tasks where local​​​‌ optima may lure agents​ away from the globally​‌ optimal solution 94.​​ At the same time​​​‌ a parallel story is​ unfolding in reinforcement learning​‌ (RL): distributed RL is​​ a sub-field where multiple​​​‌ agents solve a task​ collectively 134. Compared​‌ to the single-agent paradigm,​​ distributed RL algorithms converge​​​‌ quicker and often achieve​ superior performance. However, these​‌ algorithms have only considered​​ full connectivity. In this​​​‌ inter-disciplinary project, we presented​ a novel learning framework​‌ that augments distributed RL​​ with the notion of​​​‌ a social network structure​ and employed it to​‌ study the hypothesis from​​ human studies that partial​​​‌ connectivity performs best in​ innovation tasks.

Cultural evolution​‌ in populations of RL​​ agents. 

We implemented such​​​‌ innovation tasks using Wordcraft,​ a recently introduced RL​‌ playground inspired from the​​ Little Alchemy 2 game​​​‌ (see left of figure​ 14 for an illustration​‌ of how this task​​ works). We considered a​​​‌ wide diversity of social​ network structures: static structures​‌ that remain constant throughout​​ learning (fully-connected, ring, small-world)​​​‌ and a dynamic structure​ where the group oscillates​‌ between phases of low​​ and high connectivity (we​​​‌ illustrate this dynamic structure​ on the right of​‌ figure 14). Each​​ agent in our implementation​​​‌ employs the DQN learning​ algorithm and exchanges experiences​‌ that have the form​​ of sequences of state-action​​ combinations with its neighbors.​​​‌

(Left) Illustration of an‌ innovation task, consisting of‌​‌ an initial set of​​ elements (Earth, Water) and​​​‌ a recipe book indicating‌ which combinations create new‌​‌ elements. Upon creating a​​ new element the player​​​‌ moves up an innovation‌ level and receives a‌​‌ reward that increases monotonically​​ with levels. (Right) Dynamic​​​‌ social network structures oscillate‌ between phases of low‌​‌ connectivity, where experience sharing​​ takes place within clusters,​​​‌ and high connectivity, where‌ experiences spread between clusters.‌​‌

A‌ central conclusion of our‌​‌ empirical analysis was that​​ the dynamic social network​​​‌ structure performs best. In‌ addition to the performance‌​‌ groups achieve we measured​​ behavioral and mnemonic metrics​​​‌ such as behavioral conformity‌ and mnemonic diversity. Such‌​‌ metrics were inspired from​​ human studies and helped​​​‌ us further analyze the‌ behavior of groups. For‌​‌ example, one empirical observation​​ was that sharing experiences​​​‌ did not help the‌ group learn quicker in‌​‌ a very simple innovation​​ task; instead the fully-connected​​​‌ group was the slowest.‌ By looking at the‌​‌ diversity in the memories​​ of the agents we​​​‌ observed that the fully-connected‌ structure had the highest‌​‌ individual diversity (left of​​ figure 15 ) and​​​‌ the lowest group diversity‌ (right of figure 15‌​‌): sharing experiences with​​ others diversifies an individual's​​​‌ experiences but also homogenizes‌ the group, which is‌​‌ bad for its performance.​​

Illustration of an innovation​​​‌ task, consisting of an‌ initial set of elements‌​‌ (Earth, Water) and a​​ recipe book indicating which​​​‌ combinations create new elements.‌ Upon creating a new‌​‌ element the player moves​​ up an innovation level​​​‌ and receives a reward‌ that increases monotonically with‌​‌ levels. Dynamic social network​​ structures oscillate between phases​​​‌ of low connectivity, where‌ experience sharing takes place‌​‌ within clusters, and high​​ connectivity, where experiences spread​​​‌ between clusters.

We see the​​​‌ contribution of this project‌ as two-fold. From the‌​‌ perspective of fields studying​​ human intelligence, we have​​​‌ shown that using RL‌ algorithms as computational tool‌​‌ is a promising direction​​ towards increasing the verisimilitude​​​‌ of simulations and analyzing‌ both behavior and memory.‌​‌ From the perspective of​​​‌ RL, we have shown​ that distributed RL algorithm​‌ should move beyond the​​ fully-connected architecture and explore​​​‌ groups with dynamic topologies.​ This work is currently​‌ a preprint 136 and​​ is about to be​​​‌ submitted in PNAS. We​ open-source the code at​‌ this link.

Cultural​​ evolution in populations of​​​‌ LLM agents. 

In 2024,​ we have extended this​‌ framework with agents equipped​​ with Large Language Models​​​‌ (LLMs) playing Little Alchemy​ 2, a creative video​‌ game originally developed for​​ humans (figure 16).​​​‌ We, first, study an​ LLM in isolation and​‌ discover that it exhibits​​ both useful skills and​​​‌ crucial limitations. We, then,​ study groups of LLMs​‌ that share information related​​ to their behaviour and​​​‌ focus on the effect​ of social connectivity on​‌ collective performance. In agreement​​ with previous human and​​​‌ computational studies (including the​ one described above), we​‌ observe that groups with​​ dynamic connectivity out-compete fully-connected​​​‌ groups. Our work reveals​ opportunities and challenges for​‌ future studies of collective​​ innovation that are becoming​​​‌ increasingly relevant as Generative​ Artificial Intelligence algorithms and​‌ humans innovate alongside each​​ other. We published this​​​‌ work at the ALife​ 2024 conference 139.​‌

Studying collective innovation in​​ groups of LLMs: A)​​​‌ we experiment with Little​ Alchemy 2, a game​‌ where players combine real-world​​ items to create new​​​‌ ones. A knowledge graph​ describes the possible combinations​‌ (we only present a​​ small sub-part of the​​​‌ graph which contains 720​ items in total) B)​‌ Alice-LLM and Bob-LLM are​​ two LLMs playing the​​​‌ game together. They are​ provided with the same​‌ intro prompt, explaining the​​ rules of the game,​​​‌ and the same task​ (they start with the​‌ same set of items).​​ Alice-LLM and Bob-LLM have​​​‌ identical weights but behave​ differently because the state​‌ prompt depends on their​​ crafting history. They are​​​‌ informed about the actions​ of others through their​‌ prompt. In this paper,​​ we study how groups​​​‌ of such LLM agents​ are able to efficiently​‌ explore a knowledge graph,​​ focusing in particular on​​​‌ the effect of different​ social structures specifying with​‌ whom and when they​​ can share information

8.2.2 When LLMs‌​‌ Play the Telephone Game:​​ Cultural Attractors as Conceptual​​​‌ Tools to Evaluate LLMs‌ in Multi-turn Settings

Participants:‌​‌ Jérémy Perez [correspondant],​​ Grgur Kovač, Corentin​​​‌ Léger, Cédric Colas‌, Gaia Molinaro,‌​‌ Maxime Derex, Pierre-Yves​​ Oudeyer, Clément Moulin-Frier​​​‌.

As large language‌ models (LLMs) start interacting‌​‌ with each other and​​ generating an increasing amount​​​‌ of text online, it‌ becomes crucial to better‌​‌ understand how information is​​ transformed as it passes​​​‌ from one LLM to‌ the next. While significant‌​‌ research has examined individual​​ LLM behaviors, existing studies​​​‌ have largely overlooked the‌ collective behaviors and information‌​‌ distortions arising from iterated​​ LLM interactions. Small biases,​​​‌ negligible at the single‌ output level, risk being‌​‌ amplified in iterated interactions,​​ potentially leading the content​​​‌ to evolve towards attractor‌ states.

In this project,‌​‌ we ran a series​​ of telephone game experiments,​​​‌ applying a transmission chain‌ design borrowed from the‌​‌ human cultural evolution literature:​​ LLM agents iteratively receive,​​​‌ produce, and transmit texts‌ from the previous to‌​‌ the next agent in​​ the chain.

This figures​​​‌ depicts the method to‌ estimate the strength and‌​‌ position of theoretical attractors.​​ Each dot in this​​​‌ figure corresponds to one‌ chain, for a total‌​‌ of 100 chains (20​​ initial texts * 5​​​‌ seeds). The position of‌ a dot on the‌​‌ x-axis corresponds to the​​ value of the property​​​‌ (positivity in this‌ example) in the initial‌​‌ text, while the position​​ on the y-axis corresponds​​​‌ to the value of‌ this property of the‌​‌ text produced after 50​​ generations. We then used​​​‌ these 100 data points‌ to fit a linear‌​‌ regression predicting the relationship​​ between the initial and​​​‌ final values of the‌ property.

Our​​​‌ main contributions are:

• We‌ propose that there might‌​‌ be a gap in​​ current LLM evaluations methods​​​‌ (single-turn evaluations might not‌ be suited to assess‌​‌ the properties of multi-turn​​ interactions)
• We empirically confirm​​​‌ this hypothesis by showing‌ that multi-turn interactions indeed‌​‌ often lead to distributions​​ of text properties that​​​‌ are significantly different from‌ what is observed after‌​‌ a single interaction.
• We​​ introduce novel conceptual and​​​‌ methodological tools to fill‌ this gap, grounded in‌​‌ research in cultural evolution,​​ and in particular the​​​‌ concept of cultural attractor.‌
• We showcase the potential‌​‌ of this method by​​ applying it to compare​​​‌ the effect of different‌ tasks, of different models,‌​‌ of temperature, and of​​ fine-tuning on the properties​​​‌ of multi-turn interactions.
• We‌ find several robust effects,‌​‌ such as the fact​​ that less constrained tasks​​​‌ lead to stronger attractors,‌ that some properties posses‌​‌ stronger attractors than others,​​ and that fine-tuning can​​​‌ shift the position and‌ modify the strength of‌​‌ attractors.

The heigth of​​ the bars represent the​​​‌ position (top row) and‌ strength (bottom row) of‌​‌ theoretical attractors, for each​​ property (columns), task, and​​​‌ model. Less constrained tasks,‌ such as Continue,‌​‌ appear to produce stronger​​​‌ attractors than more constrained​ tasks, such as Rephrase​‌. Attractors appear to​​ be stronger for toxicity​​​‌ than for length.​ Finally, we can notice​‌ that the position of​​ attractors appears to vary​​​‌ between models.

These findings highlight​​ the importance of accounting​​​‌ for multi-step transmission dynamics​ and represent a first​‌ step towards a more​​ comprehensive understanding of LLM​​​‌ cultural dynamics.

This work​ was presented during a​‌ 15-minutes talk given at​​ the 2024 Cultural Evolution​​​‌ Society conference, and was​ accepted as a conference​‌ paper at the International​​ Conference on Representation Learning​​​‌ 2025 (ICLR 2025) Conference​ 144 The code is​‌ available at here.​​ We also created a​​​‌ website featuring a Data​ Explorer tool, allowing to​‌ directly inspect the texts​​ generated during our experiments.​​​‌

8.2.3 Recursive Training Loops​ in LLMs: How training​‌ data properties modulate distribution​​ shift in generated data?​​​‌

Participants: Grgur Kovač [correspondant]​, Jérémy Perez [correspondant]​‌, Remy Portelas,​​ Peter Ford Dominey,​​​‌ Pierre-Yves Oudeyer.

Large​ language models (LLMs) are​‌ increas- ingly used in​​ the creation of online​​​‌ content, creating feedback loops​ as subsequent gener- ations​‌ of models will be​​ trained on this syn-​​​‌ thetic data. Such loops​ were shown to lead​‌ to distribution shifts -​​ models misrepresenting the true​​​‌ underlying distributions of human​ data (also called model​‌ collapse). However, how hu-​​ man data properties affect​​​‌ such shifts remains poorly​ understood. In this paper,​‌ we provide the first​​ empirical examination of the​​​‌ effect of such properties​ on the outcome of​‌ recursive training. We first​​ confirm that using differ-​​​‌ ent human datasets leads​ to distribution shifts of​‌ different magnitudes. Through exhaustive​​ manipulation of dataset properties​​​‌ combined with regression analyses,​ we then identify a​‌ set of properties associated​​ with distribution shift magnitudes.​​​‌ Lexical diversity is found​ to am- plify these​‌ shifts, while semantic diversity​​ and data quality mitigate​​​‌ them. Furthermore, we find​ that these influences are​‌ highly modular: data scrapped​​ from a given internet​​​‌ domain has little influence​ on the content generated​‌ for an- other domain.​​ Finally, experiments on political​​​‌ bias reveal that human​ data properties affect whether​‌ the initial bias will​​ be amplified or re-​​​‌ duced. Overall, our results​ portray a novel view,​‌ where different parts of​​ internet may undergo different​​​‌ types of distribution shift.​

The main contributions of​‌ this work are:

• We​​ propose and experimentally confirm​​​‌ the hypothesis that different​ training datasets lead to​‌ different distribution shift dynamics,​​ motivating an investigation on​​​‌ the underlying causes.
• Through​ an extensive set of​‌ experiments (four datasets over​​ three domains), we outline​​​‌ several data properties as​ influencing distribution shift dynamics.​‌
• We reveal that these​​ influences are highly modular,​​​‌ with generated content being​ mostly influenced by human​‌ data properties from the​​ same domain.
• We find​​​‌ that distribution shifts also​ occur in terms of​‌ political lean, and that​​ the type of shift​​​‌ (bias amplification, reduction or​ inversion) depends on the​‌ political lean of the​​ human data.

Iterative chain​​ In each generation, a​​​‌ fresh base model is‌ fine-tuned on texts sampled‌​‌ from the Accumulated data​​ pool (except generation 0,​​​‌ where it's trained only‌ on human posts). The‌​‌ model generates posts, which​​ are added to the​​​‌ pool alongside some newly‌ sampled human posts.

This‌​‌ work was published as​​ a conference paper at​​​‌ the EMNLP2025 conference 49‌.

8.2.4 Intrinsic motivation‌​‌ is key to understanding​​ peer cultures

Participants: Jérémy​​​‌ Perez [correspondant], Maxime‌ Derex, Pierre-Yves Oudeyer‌​‌, Clément Moulin-Frier.​​

This paper 64 is​​​‌ a commentary to 123‌, as part of‌​‌ the call for open-peer​​ commentary on this target​​​‌ article in Behavioral and‌ Brain Sciences, 1–68. In‌​‌ the target paper, the​​ authors make an intriguing​​​‌ case that peer cultures‌ could play a key‌​‌ role in cultural adaptation​​ by generating qualitatively different​​​‌ cultural variation compared to‌ adult cultures. However, the‌​‌ mechanisms responsible for this​​ distinction remain unclear. In​​​‌ out commentary, we discuss‌ how accounting for the‌​‌ role of intrinsic motivation​​ in shaping the content​​​‌ of peer cultures may‌ help explain their evolutionary‌​‌ dynamics.

8.2.5 Cultural variation​​ and regularities in intrinsically​​​‌ motivated exploration: investigating autonomous‌ goal selection in BaYaka‌​‌ foragers and Bandongo fisher-farmers​​

Participants: Jérémy Perez [correspondant]​​​‌, Sarah Pope-Caldwell,‌ Sheina Lew-Levy, Pierre-Yves‌​‌ Oudeyer, Maxime Derex​​, Clément Moulin-Frier.​​​‌

TLDR: This study investigates‌ how recent performance and‌​‌ recent progress influence autonomous​​ goal selection in children​​​‌ and adults from two‌ cultural groups in the‌​‌ Congo Basin. All data​​ necessary for this project​​​‌ has been collected during‌ Jérémy Perez's mission in‌​‌ Congo in July-August 2025,​​ and analyses are still​​​‌ ongoing. This project was‌ made possible through a‌​‌ collaboration with Sheina Lew-Levy​​ from Durham University and​​​‌ Sarah Pope-Caldwell from Georgia‌ State University. An abstract‌​‌ for this project will​​ be submitted to the​​​‌ 2026 conference of the‌ European Human Behaviour &‌​‌ Evolution Association.

8.2.6 The cultural evolution​​ of human goals: How​​​‌ individuals generate, select, and​ transmit goals

Participants: Jérémy​‌ Perez [correspondant], Cédric​​ Colas, Gaia Molinaro​​​‌, Pierre-Yves Oudeyer,​ Maxime Derex, Clément​‌ Moulin-Frier.

This work​​ has been submitted to​​​‌ the special issue on​ Goal Dynamics in Cognition​‌ of the journal Topics​​ in Cognitive Science, and​​​‌ is currently under review.​

Abstract:

Humans pursue goals​‌ that are remarkably diverse​​ and vary over time​​​‌ and cultures. These goals​ shape which behaviors are​‌ explored, valued, and socially​​ transmitted, yet most theories​​​‌ of cultural evolution focus​ on how behaviors evolve​‌ while leaving the origins​​ of goals unexamined. We​​​‌ argue that a complete​ understanding of cultural evolution​‌ requires explaining how goals​​ themselves emerge, vary, and​​​‌ persist across generations. Building​ on studies of motivation​‌ and curiosity in cognitive​​ science and artificial intelligence,​​​‌ we introduce the notion​ of cultural autotelic agents—individuals​‌ who actively generate, select,​​ and transmit their own​​​‌ goals within social environments.​ By highlighting the cognitive​‌ and motivational mechanisms that​​ drive goal formation and​​​‌ selection, this framework extends​ existing models of cultural​‌ evolution and helps explain​​ the open-ended, self-propelling character​​​‌ of human culture.

We​ introduce the notion of​‌ cultural autotelic agents, i.e.​​ agents that combine individual​​​‌ and social learning to​ represent, generate, select, and​‌ transmit their own goals.​​ This model departs from​​​‌ the historical conceptualization as​ problem-solvers (left column). This​‌ standard perspective focuses on​​ how agents optimize behaviors​​​‌ toward goals that are​ externally imposed. This view​‌ is largely present in​​ research on individual cognition​​​‌ (top-left) and has inspired​ most experimental paradigms in​‌ cultural evolution (bottom left,​​ e.g. transmission chains). Research​​​‌ on motivation, developmental psychology,​ and developmental artificial intelligence​‌ has extended this view​​ toward the concept of​​​‌ autotelic agents, i.e.​ agents able to self-generate​‌ and purse their own​​ goals (top-right). This conceptualization​​​‌ affords a more complete​ understanding of proactively exploratory​‌ behaviors in humans, in​​ particular how their past​​​‌ behavior influence their goal​ generation and selection mechanisms.​‌ Here, we propose integrating​​ such insights from cultural​​ evolution and autotelic learning​​​‌ to introduce the concept‌ of cultural autotelic agents‌​‌ (bottom-right). Under this view,​​ agents are active in​​​‌ the generation and selection‌ of the goals they‌​‌ pursue. These goal generation​​ and selection mechanisms are​​​‌ influenced both by social‌ information and individually collected‌​‌ information. We argue that​​ this conceptualization is necessary​​​‌ to think about the‌ cultural evolutionary dynamics of‌​‌ goals.

8.2.7 Evolving Interaction Protocols​​​‌ for Open-Ended Collective Innovation‌

Participants: Akhi Mocherla,‌​‌ Jérémy Perez, Eleni​​ Nisioti, Cédric Colas​​​‌.

In exploratory domains‌ such as science, art,‌​‌ and design, progress emerges​​ not from achieving predefined​​​‌ objectives but from accumulating‌ novel and meaningful discoveries‌​‌ 140. Lab and​​ field studies of human​​​‌ collective innovation have shown‌ that a group's exploration‌​‌ and, thus, innovation abilities​​ critically depend on how​​​‌ individuals communicate with each‌ other 94, 130‌​‌, 124. For​​ example, increasing group connectivity​​​‌ speeds up innovation in‌ the short-term but reduces‌​‌ diversity within the collective,​​ negatively impacting long-term innovation.​​​‌ Partially connected groups thus‌ accumulate the most innovations‌​‌ in deceptive search spaces.​​ Computational studies have confirmed​​​‌ this in groups of‌ evolving agents 121,‌​‌ 78, reinforcement learning​​ agents 137, and​​​‌ Large Language Models (LLMs)‌ 138, highlighting the‌​‌ key role of collective​​ dynamics in engineered multi-agent​​​‌ systems. Despite this, systematic‌ approaches for optimising how‌​‌ groups interact remain underdeveloped.​​ In this work, we​​​‌ propose an approach for‌ designing interaction protocols (IPs)‌​‌ that govern who communicates​​ with whom, what is​​​‌ communicated, and when. Similarly‌ to past computational studies‌​‌ 137, 138,​​ we use the text-based​​​‌ game Little Alchemy 2‌ (LA2) as a test-bed‌​‌ of collective innovation. To​​​‌ explore the IP space​ systematically, we employ a​‌ Quality-Diversity (QD) algorithm 152​​ discovering repertoires with high​​​‌ performance and behavioral diversity.​ We maintain an archive​‌ of IPs, each evaluated​​ via multiple trials. Similarly​​​‌ to previous works 100​ , our approach follows​‌ the Novelty Search with​​ Local Competition (NS-LC) paradigm​​​‌ and employs LLMs within​ QD for solution generation​‌ and novelty estimation.

This​​ work has been presented​​​‌ as a poster at​ the 2025 workshop on​‌ Intrinsically Motivated Open-ended Learning​​ (IMOL 2025).

Overview of​​​‌ the framework used to​ evolve Interaction Protocols for​‌ groups of agents playing​​ the Little Alchemy 2​​​‌ game. The system iteratively​ generates new IPs using​‌ a language model (LLM),​​ evaluates their performance, and​​​‌ maintains an archive of​ candidate solutions. Candidate IPs​‌ are debugged and tested,​​ then evaluated for fitness​​​‌ and novelty relative to​ archived solutions. The archive​‌ is updated based on​​ fitness, and novel or​​​‌ improved protocols are used​ to guide further LLM​‌ generations, enabling continual improvement​​ and diversity in discovered​​​‌ solutions. Comparison of the​ performance of an evolved​‌ IP to dynamic and​​ fully-connected IPs from past​​​‌ studies.

8.2.8 Inferring the​​ Phylogeny of Large Language​​​‌ Models and Predicting their​ Performances in Benchmarks

Participants:​‌ Nicolas Yax [correspondant],​​ Pierre-Yves Oudeyer, Stefano​​​‌ Palminteri.

In recent​ month the number of​‌ Large Language Models (LLMs)​​ released has never been​​​‌ that high. On one​ hand, multiple private companies​‌ such as OPENAI, Claude,​​ Google, Mistral, etc. are​​​‌ making cutting-edge models that​ have a lot of​‌ visibility in our modern​​ society and science. However,​​​‌ as the number of​ LLMs is raising, the​‌ training methods are becoming​​ more secretive making the​​​‌ field increasingly obscur to​ science. On the other​‌ hand, everyday, a few​​ hundreds of open-access language​​​‌ models are uploaded on​ the hugging face hub​‌ which is far too​​ much to keep track​​​‌ of the evolution of​ LLMs in the field.​‌ Knowing that not all​​ of these open models​​​‌ are perfectly transparent about​ the training methods and​‌ only very few of​​ them are benchmarked (due​​​‌ to the high cost​ of benchmarking) there is​‌ an increasing need for​​ methods to help keep​​​‌ track of the progress​ and evolution of these​‌ models in the field.​​

We developped an algorithm,​​​‌ named PhyloLM, inspired from​ phylogenetics to compute evolutionary​‌ trees in LLMs. We​​ show this method efficient​​ in reconstructing the evolutionary​​​‌ history of LLMs within‌ families 22, in‌​‌ discriminating the different families,​​ and also in finding​​​‌ similarities between these families.‌ Additionaly, the genetic information‌​‌ can be used to​​ predict LLM capabilities like​​​‌ benchmark scores showing a‌ very significant correlation between‌​‌ predicted and true scores.​​ These advances could be​​​‌ instrumental in our way‌ to navigate the field‌​‌ of LLMs by making​​ the world of LLM​​​‌ more transparent at a‌ very low cost. This‌​‌ was published at ICLR​​ 2025.

Phylogenetic tree​​​‌ reconstruction. On the left‌ it is shown the‌​‌ ground truth concerning the​​ relation of some LLMs​​​‌ of the Mistral family.‌ Right is the reconstruction‌​‌ from the phylogenetic algorithm​​ for the five latest​​​‌ models of this family‌ ("leaves" of the phylogenetic‌​‌ tree) on which we​​ run PhyloLM. On the​​​‌ right, it is shown‌ the reconstructed phylogenetic tree‌​‌ PhyloLM on the 5​​ "leafs" models. The numerical​​​‌ labels (0:3) map the‌ true common ancestors (on‌​‌ the right, "ground truth")​​ to the inferred ones​​​‌ (on the left, "reconstructed").‌ It can be seen‌​‌ that the true and​​ the reconstructed trees are​​​‌ topologically equivalent

8.3.1 Research perspective: The‌ Ecology of Open-Ended skill‌​‌ Acquisition

Participants: Clément Moulin-Frier​​ [correspondant], Eleni Nisioti​​​‌, Pierre-Yves Oudeyer.‌

An intriguing feature of‌​‌ the human species is​​ our ability to continuously​​​‌ invent new problems and‌ to proactively acquiring new‌​‌ skills in order to​​ solve them: what is​​​‌ called Open-Ended Skill Acquisition‌ (OESA). Understanding the mechanisms‌​‌ underlying OESA is an​​ important scientific challenge in​​​‌ both cognitive science (e.g.‌ by studying infant cognitive‌​‌ development) and in artificial​​ intelligence (aiming at computational​​​‌ architectures capable of open-ended‌ learning). Both fields, however,‌​‌ mostly focus on cognitive​​ and social mechanisms at​​​‌ the scale of an‌ individual’s life. It is‌​‌ rarely acknowledged that OESA,​​ an ability that is​​​‌ fundamentally related to the‌ characteristics of human intelligence,‌​‌ has been necessarily shaped​​ by ecological, evolutionary and​​​‌ cultural mechanisms interacting at‌ multiple spatiotemporal scales.

The‌​‌ ORIGINS framework identifies central​​ components (boxes) and their​​​‌ interactions (arrows) driving Open-Ended‌ Skill Acquisition, both in‌​‌ terms of its evolution​​ from environmental complexity (roughly:​​​‌ left to right arrows)‌ as well its open-ended‌​‌ aspect through feedback mechanisms​​ (right to left arrows).​​​‌ The employed terminology reflects‌ a diversity of mechanisms‌​‌ considered in both Artificial​​​‌ Intelligence and Human Behavioral​ Ecology.

We have​ recently initiated a new​‌ research direction aiming at​​ understanding, modeling and simulating​​​‌ the dynamics of OESA​ in artificial systems, grounded​‌ in theories studying its​​ eco-evolutionary bases in the​​​‌ human species. For this​ aim, we have proposed​‌ a conceptual framework, called​​ ORIGINS (illustrated Fig. 23​​​‌ and developed in 131​), expressing the complex​‌ interactions between environmental, adaptive,​​ multi-agent and cultural dynamics.​​​‌ This framework raises three​ main research questions:

• What​‌ are the ecological conditions​​ favoring the evolution of​​​‌ autotelic agents?
• How to​ bootstrap the formation of​‌ a cultural repertoire in​​ populations of adaptive agents?​​​‌
• What is the role​ of cultural feedback effects​‌ in the open-ended dynamics​​ of human skill acquisition?​​​‌

The contributions described below​ are addressing some aspects​‌ of these research questions.​​ Note that there might​​​‌ be a thematic overlap​ between the two last​‌ research questions outlined above​​ and the previous section​​​‌ on Models of Cultural​ Evolution 8.2, where​‌ we also present related​​ results.

8.3.2 Eco-evolutionary Dynamics​​​‌ of Non-episodic Neuroevolution in​ Large Multi-agent Environments

Participants:​‌ Gautier Hamon [correspondant],​​ Eleni Nisioti, Clément​​​‌ Moulin-Frier.

This work​ was published in 2023​‌ but we keep it​​ in this report as​​​‌ it introduces a general​ computational framework, called non-episodic​‌ neuroevolution, that forms the​​ basis of the two​​​‌ next contributions.

This contribution​ focuses on eco-evolutionary dynamics​‌ where "organisms are not​​ solely products but, by​​​‌ modifying their niche and​ therefore its associated fitness​‌ landscape, are also causes​​ of evolution" 120.​​​‌ The main objective of​ this paper is to​‌ propose a method for​​ studying large-scale eco-evolutionary dynamics​​​‌ in agent-based simulations with​ a reasonable level of​‌ biological and ecological plausibility.​​ For this aim, we​​​‌ implement a system with​ the following properties (see​‌ Fig. 24 for illustration):​​

• Non-episodic simulation environment with​​​‌ complex intrinsic dynamics.​ We model our environment​‌ after common-pool resource (CPR)​​ appropriation problems, where a​​​‌ group of agents competes​ for finite resources. We​‌ extend an existing environment​​ of CPR appropriation  145​​​‌ with the presence of​ multiple niches, where resources​‌ regrow proportionally to the​​ density of nearby resources​​​‌ at different rates in​ different regions of the​‌ environment (Fig 24).​​ We prevent any environment​​​‌ or population reset during​ a whole simulation run,​‌ enabling coupled environmental and​​ population dynamics leading to​​​‌ complex eco-evolutionary feedback effects.​
• Continuous neuroevolution in a​‌ large, size-varying agent population​​ The environment contains thousands​​​‌ of agents, each controlled​ by a neural network​‌ whose weights are optimized​​ using neuroevolution 161
• Physiology-driven​​​‌ death and reproduction There​ is no notion of​‌ rewards, agents are instead​​ equipped with a physiological​​ system modulating their energy​​​‌ level according to the‌ resources they consume, in‌​‌ a non-linear way. At​​ the evolutionary scale, agents​​​‌ reproduce as long as‌ they are able to‌​‌ maintain their energy level​​ within a reasonable range​​​‌ and die if this‌ level goes below a‌​‌ minimum threshold. This is​​ departure from the notion​​​‌ of fitness-based selection and‌ more in line with‌​‌ a minimal criterion selection​​ 76. Note that​​​‌ the population size can‌ vary with time.

Our‌​‌ simulation environment (Left) is​​ an extension of the​​​‌ Common Pool Resource (CPR)‌ environment 145, 122‌​‌ : a two-dimensional grid-world​​ where some cells contain​​​‌ resources (in green) that‌ the agents (in black)‌​‌ can collect. Resources grow​​ depending on the presence​​​‌ of other resources around‌ them (local growth, Middle)‌​‌ with an additional very​​ sparse spontaneous growth, which​​​‌ means that over-consumption may‌ lead to their local‌​‌ depletion. We introduce a​​ latitudinal model of resource​​​‌ regrowth. We prevent any‌ environment and population reset‌​‌ during a whole simulation,​​ enabling continual eco-evolutionary dynamics​​​‌ to take place. Each‌ agent may reproduce or‌​‌ die according to a​​ physiological model modulating its​​​‌ energy level as a‌ function of life time‌​‌ and resource consumption (Top-Right).​​ The population size varies​​​‌ during the simulation according‌ to the current amount‌​‌ of available resources and​​ the current ability of​​​‌ agents to collect them.‌ Evolution occurs through the‌​‌ mutation of a parent's​​ network weights when it​​​‌ produces an offspring.

In addition‌ to experiments conducted in‌​‌ the large environment presented,​​ we also conduct experiments​​​‌ in "lab environment" (as‌ opposed to the "natural‌​‌ environment") to isolate the​​ study of certain behavior​​​‌ (which are often intertwined‌ with a lot of‌​‌ dynamics in the natural​​ environment).

One interesting results​​​‌ of these simulation is‌ the emergence of sustainable‌​‌ foragers which as shown​​ in lab environment Fig.​​​‌25 tends to not‌ overconsume when there is‌​‌ enough resource in their​​ neighbourhood. This allows to​​​‌ keep a certain amount‌ of resource to spread‌​‌ which is therefore beneficial​​​‌ for their future survival​ as well as the​‌ survival of their offspring.​​ (as there is no​​​‌ reset of the environment)​

Greediness of a sustainable​‌ forager agent across evaluation​​ environments that differ in​​​‌ the amount of resources.​ Sustainable agents are far​‌ less greedy in environments​​ where there is a​​​‌ certain amount of resources​ available. This strategy allows​‌ to keep resources so​​ that they spread and​​​‌ avoid overdepletion of resources.​

This​‌ work was published at​​ the Genetic and Evolutionary​​​‌ Computation Conference (GECCO) 2023.​ The computational framework it​‌ introduced led to the​​ two next recent contributions.​​​‌

8.3.3 Emergent kin selection​ of altruistic feeding via​‌ non-episodic neuroevolution

Participants: Max​​ Taylor-Davies, Gautier Hamon​​​‌, Timothe Boulet,​ Clément Moulin-Frier [correspondant].​‌

This work extends the​​ project presented in previous​​​‌ contribution Sec.8.3.2.​ It is the result​‌ from the visit in​​ the team of Max​​​‌ Taylor-Davies doing his PhD​ at School of Informatics,​‌ University of Edinburgh, Scotland.​​ It has been accepted​​​‌ at the EvoStar conference​ the International Conference on​‌ the Applications of Evolutionary​​ Computation (Part of EvoStar)​​​‌ 55.

At first​ glance, it seems difficult​‌ to square the phenomenon​​ of purely altruistic behaviour​​​‌ (acts which confer a​ benefit to the recipient​‌ at a cost to​​ the actor) with the​​​‌ basic principle of natural​ selection: how can a​‌ gene be selected for​​ when it decreases, rather​​​‌ than increases, the fitness​ of its host? One​‌ plausible account can be​​ made through the theory​​​‌ of inclusive fitness. Key​ to this theory is​‌ the recognition that individual​​ organisms within a social​​​‌ environment are not isolated​ from their conspecifics in​‌ terms of fitness. Whether​​ a given gene is​​​‌ selected for is thus​ determined by its effect(s)​‌ on the fitness of​​ any bearers of copies​​​‌ of that gene. Under​ this view, we can​‌ think of an altruistic​​ act as an exchange​​​‌ of fitness from one​ agent to another. If​‌ the exchange is positive-sum​​ and both sides are​​​‌ bearers of the gene​ in question, then from​‌ the gene's perspective the​​ behaviour confers a fitness​​​‌ benefit–even while it decreases​ the fitness of the​‌ acting individual.

Kin selection​​ theory 160 has proven​​​‌ to be a popular​ and widely accepted account​‌ of how altruistic behaviour​​ can evolve under natural​​​‌ selection. Hamilton's rule, first​ published in 1964 107​‌, 108, has​​ since been experimentally validated​​​‌ across a range of​ different species and social​‌ behaviours. In contrast to​​ this large body of​​​‌ work in natural populations,​ however, there has been​‌ relatively little study of​​ kin selection in silico​​​‌. In the current​ work, we offer what​‌ is to our knowledge​​ the first demonstration of​​ kin selection emerging naturally​​​‌ within a population of‌ agents undergoing continuous neuroevolution.‌​‌ Specifically, we find that​​ zero-sum transfer of resources​​​‌ from parents to their‌ infant offspring evolves through‌​‌ kin selection in environments​​ where it is hard​​​‌ for offspring to survive‌ alone. In an additional‌​‌ experiment, we show that​​ kin selection in our​​​‌ simulations relies on a‌ combination of kin recognition‌​‌ and population viscosity. We​​ believe that our work​​​‌ may contribute to the‌ understanding of kin selection‌​‌ in minimal evolutionary systems,​​ without explicit notions of​​​‌ genes and fitness maximisation.‌

The relationship between the‌​‌ estimated benefit to infants​​ of being fed and​​​‌ both the amount and‌ selectivity of feeding observed,‌​‌ shown separately for each​​ of the three experimental​​​‌ parameters we varied (and‌ combined in the rightmost‌​‌ column). Each scatterplot point​​ represents a single 500k-timestep​​​‌ simulation run (with values‌ averaged over the final‌​‌ 50k timesteps); regression lines​​ (with 95% confidence intervals)​​​‌ are shown in green.‌ Note that the $\mathrm{y‌‌}$-axis shows $log$(measure)​​ for both amount and​​​‌ selectivity.

This paper was‌ accepted at The International‌​‌ Conference on the Applications​​ of Evolutionary Computation (EvoAPPS)​​​‌ 2025 (part of EvoStar).‌

8.3.4 Evolving large populations‌​‌ of adaptive neural agents​​ in ecologically plausible environments​​​‌

Participants: Timothé Boulet [correspondant]‌, Gautier Hamon,‌​‌ Clément Moulin-Frier.

This​​ work continues the project​​​‌ presented in the previous‌ paragraph, with a focus‌​‌ on the ability of​​ agents to develop adaptability​​​‌ behaviors. Specifically, we extend‌ the framework by adding‌​‌ fruits, a spatially variable​​ ressource, and a memory​​​‌ of the values of‌ each type of fruits‌​‌ for the agents. The​​ goal is to observe​​​‌ whether the agents manage‌ to exploit the knowledge‌​‌ of the fruits values​​ to decide which fruit​​​‌ exploit.

Results : the‌ agent were able to‌​‌ exploit the fruit value​​ information to optimize their​​​‌ behavior. There were also‌ some results that we‌​‌ were not necessarily expecting​​ and that comes from​​​‌ our choice of model.‌ Notably, it seems the‌​‌ agents choice for exploiting​​ a cluster is heavily​​​‌ influenced by social criteria‌ (the number of agents‌​‌ already exploiting it) and​​ cultural criteria (whether the​​​‌ cluster is empty or‌ full of fruits). This‌​‌ effect exceeds the adapatability​​ effect in the latest​​​‌ stages of the simulations.‌

8.4.1 DevCur Project: studying​​​‌ the co-development of curiosity,‌ metacognition and agency in‌​‌ adolescents

Participants: Julien Rosenberger​​ [correspondent], Pierre-Yves Oudeyer​​​‌, Hélène Sauzéon.‌

Under the scope of‌​‌ the DevCur project, the​​​‌ PhD of Julien Rosenberg​ was started on the​‌ following topic: “How curiosity​​ enhances learning across childhood​​​‌ and adolescence: Models and​ experimentation of the role​‌ of metacognition and agency”.​​ After exploring the literature,​​​‌ a specific project was​ settled that aims to​‌ compare the personality constructs​​ around the intellect. The​​​‌ investigated personality constructs are​ metacognitive skills (eg, 148​‌), curiosity traits (eg,​​ 115), sense of​​​‌ agency (eg, 162)​ and intellectual humility (eg,​‌ 73). Intellectual humility​​ is about correctly setting​​​‌ one’s cognitive limitations 169​. The self-report scale​‌ of Alfano et al​​ 73 hinges intellectual humility​​​‌ on other intellectual traits:​ open-mindedness (recognizing one’s cognitive​‌ limitations and having appetite​​ for knowledge without concerns​​​‌ for social status), intellectual​ modesty (having low concern​‌ for being deemed smart),​​ engagement (being able to​​​‌ confront oneself to what​ one doesn’t understand or​‌ is different from one’s​​ perspective) and corrigibility (being​​​‌ emotionally stable when one​ is intellectually challenged). The​‌ labels are slightly odd​​ but emphasize the diversity​​​‌ of intellectual traits one​ could consider in learning.​‌

A first axis is​​ to understand the organization​​​‌ of those constructs. For​ instance, intellectual humility has​‌ been linked to greater​​ general knowledge and a​​​‌ tendency to underestimate one’s​ cognitive ability 119.​‌ Those dependent variables are​​ also respectively related to​​​‌ curiosity trait 167 and​ low self-esteem and low​‌ metacognition 165. A​​ second axis is to​​​‌ obtain behavioral markers of​ those constructs. This need​‌ for situationally-bounded measures is​​ crucial for intellectual humility​​​‌ 164. It is​ currently measured through self-reports​‌ or other-reports. Yet, self-reports​​ pose an issue because​​​‌ being intellectually humble is​ socially desirable 119,​‌ requires some recall to​​ form that self-referenced attribute​​​‌ and faces the paradox​ of self-attribution (ie some​‌ humility is required to​​ say if one is​​​‌ humble). The other reports​ are alternatively resource intensive​‌ and brings other factors​​ (context, relationship…).

8.5.1 Investigating the​‌ use of LLM in​​ middle school. 

Participants: Pierre-Yves​​​‌ Oudeyer, Hélène Sauzéon​ [correspondant], Rania Abdelghani​‌.

ChatGPT, one of​​ the most widely used​​​‌ generative AI (LLM) tools,​ has made accessing mass​‌ and personalized information easy​​ and straightforward, even for​​​‌ users without expertise in​ AI. More particularly, recent​‌ reports indicate that the​​ majority of surveyed students​​​‌ aged nine and older​ have already used this​‌ tool for school-related tasks.​​ However, while we know​​​‌ that students are using​ ChatGPT, there is limited​‌ understanding of how they​​ use it and its​​​‌ effects on their learning​ processes and outcomes, particularly​‌ among middle and high​​ school students and in​​​‌ subjects outside programming.

Investigating​ these patterns of use​‌ is a critical step​​ toward identifying the necessary​​​‌ educational interventions to mitigate​ risks associated with misuse​‌ or harmful interactions with​​ ChatGPT, which are particularly​​​‌ likely among non-expert users.​ To address this, we​‌ recruited 63 students aged​​ 14 to 15 and​​​‌ asked them to solve​ science problems using ChatGPT.​‌ We examined their prompt​​ choices, evaluations of ChatGPT's​​ responses, and final problem-solving​​​‌ outcomes. Overall, our results‌ indicate that students are‌​‌ still inefficient users of​​ AI tools such as​​​‌ ChatGPT and are vulnerable‌ to incorporating its misinformation,‌​‌ even when they report​​ high domain knowledge and​​​‌ previous experience with generative‌ AI. This highlights potential‌​‌ misconceptions about these tools’​​ capabilities and the skills​​​‌ required to use them‌ effectively. Furthermore, domain knowledge‌​‌ alone appears insufficient to​​ shield students from adopting​​​‌ misinformation generated by ChatGPT.‌ Implementing formal educational interventions‌​‌ to correct these misconceptions​​ and train students for​​​‌ informed usage thus seems‌ both timely and essential,‌​‌ given the growing reliance​​ on generative AI tools​​​‌ in education. On the‌ longer term, fostering metacognitive‌​‌ skills may further promote​​ responsible and effective use​​​‌ of such tools (paper‌ in preparation)

8.5.2 Study‌​‌ impact of a pedagogical​​ intervention on GenAI in​​​‌ middle school students

Participants:‌ Pierre-Yves Oudeyer, Hélène‌​‌ Sauzeon, Olivier Clerc​​ [correspondant], Chloé Desvaux​​​‌, Rania Abdelghani,‌ Eliott Poisson, Kan‌​‌ Yao, Didier Roy​​.

Task 

Students​​ had to solve six​​​‌ middle-school science problems using‌ ChatGPT (or a similar‌​‌ system such as DuckDuckAI).​​ Each problem included a​​​‌ statement, an image, a‌ question, and a suggested‌​‌ prompt. Students could choose​​ to use or ignore​​​‌ the suggested prompt. Two‌ types of prompts were‌​‌ provided: valid prompts (clear​​ context and precise instructions)​​​‌ and invalid prompts (insufficient‌ or vague context).

Schematic‌​‌ representation of the science​​ exercise proposed to the​​​‌ children. The experimental task‌ consisted of six science‌​‌ exercises to be completed​​ within 90 minutes. Exercises​​​‌ were provided on paper‌ sheets to prevent students‌​‌ from directly copying and​​ pasting the task description​​​‌ and accompanying image into‌ the chatbot interface.

Main Results. 

Overall, students​‌ who benefited from the​​ pedagogical intervention achieved higher​​​‌ performance than those in​ the control group. In​‌ quantitative terms, the mean​​ score (out of 20)​​​‌ was approximately 10.3 in​ the control group—comparable to​‌ the 2024 pilot study—and​​ approximately 11.4 in the​​​‌ experimental group. This difference​ is statistically significant (p​‌ < .05). The experimental​​ group not only obtained​​​‌ significantly higher scores, but​ also demonstrated more strategic​‌ use of the AI​​ system. In particular, they​​​‌ rejected invalid prompts more​ frequently and were more​‌ likely to reformulate or​​ refine their queries when​​​‌ the initial answer was​ unsatisfactory. Moreover, formulating their​‌ own prompts tended to​​ maintain or improve performance,​​​‌ even in cases where​ the suggested prompt was​‌ already valid. Importantly, self-reported​​ prior knowledge about AI​​​‌ was not associated with​ better performance, suggesting that​‌ explicit instruction and practice​​ played a more decisive​​​‌ role than familiarity with​ the technology alone.

Workshop​‌ effects on performance and​​ prompt acceptance

8.5.3 LLM4Humanities:​ An Open-Source Toolkit for​‌ LLM-Assisted Qualitative Research

Participants:​​ Olivier Clerc [correspondant],​​​‌ Grgur Kovač, Chloé​ Desvaux, Gaia Molinaro​‌, Pierre-Yves Oudeyer.​​

8.5.4‌​‌ GAIMHE: Generative AI and​​ Hybrid Models for Education​​​‌

Participants: Pierre-Yves Oudeyer,‌ Olivier Clerc, Hélène‌​‌ Sauzéon, EvidenceB .​​

8.6.1 New digital approaches​ for studying curiosity-driven learning​‌

Participants: Hélène Sauzeon [correspondant]​​, Pierre-Yves Oudeyer [correspondant]​​​‌, Rania Abdelghani,​ Mehdi Alaimi, Fabien​‌ Lotte, Aurélien appriou​​, Myra Fernandes,​​​‌ Edith Law, Yadurshana​ Sivashankar.

As curiosity​‌ is a recent research​​ topic, we studied some​​​‌ basic mechanisms of curiosity-based​ learning, thanks to three​‌ studies have been completed.​​

The first one regards​​​‌ a new interactive educational​ application to foster curiosity-driven​‌ question-asking in children. Determined​​ to improve children’s curiosity,​​​‌ we developed a new​ interactive system aiming to​‌ foster curiosity-related question-asking from​​ texts and their perception​​​‌ of curiosity. To assess​ its efficiency, we conducted​‌ a study with 95​​ fifth grade students of​​​‌ Bordeaux elementary schools. Two​ types of interventions were​‌ designed, one trying to​​ focus children on the​​​‌ construction of low-level question​ (i.e. convergent) and one​‌ focusing them on high-level​​ questions (i.e. divergent) with​​​‌ the help of prompts​ or questions starters models.​‌ We observed that both​​ interventions increased the number​​​‌ of divergent questions, the​ question fluency performance, while​‌ they did not significantly​​ improve the curiosity perception​​​‌ despite high intrinsic motivation​ scores they have elicited​‌ in children. The curiosity-trait​​ score positively impacted the​​​‌ divergent question score under​ divergent condition, but not​‌ under convergent condition. The​​ overall results supported the​​​‌ efficiency and usefulness of​ digital applications for fostering​‌ children’s curiosity that we​​ need to explore further.​​​‌ The overall results are​ published in CHI'20 72​‌. In parallel to​​ these first experimental works,​​​‌ we wrote this year​ a review of the​‌ existing works on the​​ subject 80.

The​​​‌ second study investigates the​ neurophysiological underpinnings of curiosity​‌ and the opportunities of​​ their use for Brain-computer​​​‌ interactions 74. Understanding​ the neurophysiological mechanisms underlying​‌ curiosity and therefore being​​ able to identify the​​​‌ curiosity level of a​ person, would provide useful​‌ information for researchers and​​ designers in numerous fields​​​‌ such as neuroscience, psychology,​ and computer science. A​‌ first step to uncovering​​ the neural correlates of​​​‌ curiosity is to collect​ neurophysiological signals during states​‌ of curiosity, in order​​ to develop signal processing​​​‌ and machine learning (ML)​ tools to recognize the​‌ curious states from the​​ non-curious ones. Thus, we​​​‌ ran an experiment in​ which we used electroencephalography​‌ (EEG) to measure the​​ brain activity of participants​​​‌ as they were induced​ into states of curiosity,​‌ using trivia question and​​ answer chains. We used​​ two ML algorithms, i.e.​​​‌ Filter Bank Common Spatial‌ Pattern (FBCSP) coupled with‌​‌ a Linear Discriminant Algorithm​​ (LDA), as well as​​​‌ a Filter Bank Tangent‌ Space Classifier (FBTSC), to‌​‌ classify the curious EEG​​ signals from the non-curious​​​‌ ones. Global results indicate‌ that both algorithms obtained‌​‌ better performances in the​​ 3-to-5s time windows, suggesting​​​‌ an optimal time window‌ length of 4 seconds‌​‌ to go towards curiosity​​ states estimation based on​​​‌ EEG signals. These results‌ have been published 74‌​‌.

Thanks to Virtual​​ reality device, a third​​​‌ study investigates the role‌ of intrinsic motivation in‌​‌ spatial learning in children​​ 159. In this​​​‌ study, the state curiosity‌ is manipulated as a‌​‌ preference for a level​​ of uncertainty during the​​​‌ exploration of new virtual‌ environments. To this end,‌​‌ a series of virtual​​ environments have been created​​​‌ and is presented to‌ children. During encoding, participants‌​‌ explore routes in environments​​ according the three levels​​​‌ of uncertainty (low, medium,‌ and high), thanks to‌​‌ a virtual reality headset​​ and controllers and, are​​​‌ later asked to retrace‌ their travelled routes. The‌​‌ exploration area and the​​ wayfinding. ie the route​​​‌ overlap between encoding and‌ retrieval phase, (an indicator‌​‌ of spatial memory accuracy)​​ are measured. Neuropsychological tests​​​‌ are also performed. The‌ results showed that there‌​‌ are better performances under​​ the medium uncertainty condition​​​‌ in terms of exploration‌ area and wayfinding score.‌​‌ These first results supports​​ the idea that curiosity​​​‌ states are a learning‌ booster. In Sivashankar et‌​‌ al. study, 10-year-old children​​ (20 females; 22 males)​​​‌ with low to high‌ trait curiosity actively explored‌​‌ virtual environments 29 containing​​ varying levels of uncertainty​​​‌ (low, medium, high) (Fig.‌ 30), after which‌​‌ memory for the route​​ travelled was assessed 159​​​‌.

First-person view and‌ bird’s-eye view of the‌​‌ three styles of virtual​​ environments. Participants only experienced​​​‌ the environments from a‌ first-person perspective.

From left​​​‌ to right: Condition 1‌ with Low Uncertainty (1‌​‌ character); Condition 2 with​​ Medium Uncertainty (3 characters);​​​‌ and Condition 3 with‌ High uncertainty (7 characters)‌​‌

As trait curiosity‌​‌ increased (31),​​ so did memory performance​​​‌ in the high uncertainty‌ condition, suggesting that children‌​‌ with high levels of​​ curiosity can better recruit​​​‌ cognitive resources within such‌ environments. Children with high‌​‌ compared to low curiosity​​ also had higher feelings​​​‌ of presence during the‌ immersive experience. Importantly, in‌​‌ environments with medium uncertainty,​​ children with low trait​​​‌ curiosity were able to‌ perform as well as‌​‌ those with high curiosity.​​ Results suggest that individual​​​‌ differences in trait curiosity‌ influences route memory in‌​‌ environments with varying levels​​ of uncertainty.

Route Memory​​​‌ Score (black circles) and‌ Intrinsic Motivation Score (white‌​‌ circles) in Low-and High-curiosity​​​‌ Groups as a Function​ of the Three Uncertainty​‌ Conditions (Low, Medium and​​ High)

8.6.2​​ Fostering curiosity and metacognition​​​‌ in classrooms

Participants: Pierre-Yves​ Oudeyer, Hélène Sauzéon​‌ [correspondant], Rania Abdelghani​​, Chloé Desvaux.​​​‌

Promoting curiosity by supporting​ divergent thniking

Previous work​‌ aimed to propose new​​ educational technologies driven by​​​‌ epistemic curiosity. A central​ question of this work​‌ was to specify the​​ impact of self-questioning aroused​​​‌ by states of curiosity​ (i.e., the identification of​‌ knowledge gaps and formulation​​ of learning goals) on​​​‌ student performance. To this​ end, a web platform​‌ called "Kids Ask" was​​ designed, developed, and tested​​​‌ in primary schools. The​ tool offered an interaction​‌ with a conversational agent​​ that trained children's abilities​​​‌ to generate curiosity-driven questions​ and use these questions​‌ to explore a learning​​ environment and acquire new​​​‌ knowledge. Results from this​ study suggested that the​‌ configuration helped enhance children's​​ questioning and exploratory behaviors;​​​‌ they also showed that​ learning progress differences in​‌ children can be explained​​ by differences in their​​​‌ curiosity-driven behaviors 69.​

Illustration of KidsAsk application​‌ interface

The ability to formulate​​​‌ curiosity-driven questions (i.e.​, new learning goals)​‌ likely relies upon divergent​​ thinking mechanisms, as suggested​​​‌ by literature highlighting links​ between curiosity and creativity​‌ 117156. In​​ this regard, a novel​​​‌ version of the Kids​ Ask training was proposed​‌ and tested in a​​ field study involving a​​​‌ total of 130 children​ aged 9 to 11​‌ years. These experiments aimed​​ to further assess the​​​‌ interplay between curiosity and​ creativity in question-asking behaviors.​‌ Drawing from creativity literature,​​ we examined the process​​​‌ of question formulation through​ associative thinking involved in​‌ creativity. To do so,​​ the conversational agent's behavior​​​‌ in "Kids Ask" was​ modified to prompt children​‌ to identify important keywords​​ from a text, then​​​‌ generate free associations based​ on their prior knowledge.​‌ Given the intricate interplay​​ between curiosity and creativity,​​​‌ it was hypothesized that​ this associative guidance would​‌ further enhance children's ability​​ to formulate divergent, curiosity-driven​​​‌ questions (as shown in​ figure 33)

Screen​‌ shot of the associative​​ method of prompting in​​​‌ Kids Ask. Children start​ off by reading a​‌ text containing highlighted keywords.​​ They are prompted by​​​‌ the conversational agent to​ choose one from the​‌ list and make a​​ free association with it,​​​‌ based on prior knowledge.​ They are to use​‌ one of or both​​ words to ask a​​​‌ divergent question

Promoting curiosity and​‌ metacognition in authentic settings​​

Curiosity-driven learning is crucial​​​‌ for academic achievement and​ autonomous learning, yet remains​‌ scarce in primary classrooms.​​ Building on our previous​​ work with the IGSA​​​‌ framework (Identify-Guess-Seek-Assess) introduced in‌ 68, we developed‌​‌ a training paradigm that​​ teaches curiosity-driven learning through​​​‌ metacognitive skills training. This‌ approach leverages Murayama's framework‌​‌ 133 by personifying the​​ four basic metacognitive skills​​​‌ as animated characters: the‌ referee (identify knowledge gaps),‌​‌ the detective (formulate predictions),​​ the explorer (seek information),​​​‌ and the second referee‌ (assess information quality).

Curiosity-driven‌​‌ learning framework and link​​ with the metacognitive skills​​​‌ we propose to train‌ as facilitators during our‌​‌ IGSA-based intervention

The two-part intervention combined‌ declarative knowledge about curiosity‌​‌ and metacognition with procedural​​ training of the four​​​‌ metacognitive strategies. The first‌ step consisted of animated‌​‌ videos explaining key concepts​​ related to curiosity, metacognition,​​​‌ and the four skills‌ through 2D characters. The‌​‌ second step involved the​​ "Kids Reflect" web-based platform,​​​‌ where conversational agents with‌ the same appearance and‌​‌ roles as the video​​ characters prompted children to​​​‌ use these skills appropriately‌ during reading-comprehension tasks (see‌​‌ figure below).

Screenshot of​​ the ”Kids Reflect”” platform​​​‌ during the training, given‌ one text

Our earlier pilot‌​‌ studies with small classroom​​ samples demonstrated the accessibility​​​‌ and positive impact of‌ this training on metacognitive‌​‌ efficiency, curiosity-driven question-asking, and​​ learning outcomes. These promising​​​‌ initial results motivated a‌ larger-scale validation study to‌​‌ assess both the intervention's​​ effectiveness and its scalability​​​‌ in authentic educational settings.‌

Main findings

Results‌ demonstrated that intervention groups‌​‌ significantly improved their divergent​​ question-asking abilities and developed​​​‌ more positive perceptions of‌ curiosity compared to the‌​‌ control group. Importantly, this​​ was the case in​​​‌ the ecological setting of‌ classrooms where teachers managed‌​‌ the intervention themselves, but​​ also with a lighter​​​‌ easy-to-implement version of the‌ training (see figure below).‌​‌

Post-interventions results of question-asking​​ abilities of children in​​​‌ each condition

However, nuanced​​​‌ findings emerged regarding teacher‌ delivery conditions. These groups‌​‌ showed lower performance during​​ the intervention and poorer​​​‌ learning outcomes, alongside higher‌ cognitive load, compared to‌​‌ researcher-led groups. This suggests​​​‌ that while the intervention​ can be effectively scaled​‌ to teacher-led implementations, some​​ avenues for improvement have​​​‌ been identified. This point​ was further informed by​‌ qualitative interviews conducted with​​ volunteered teachers who were​​​‌ animators in the study.​ Teachers rated the intervention​‌ highly on acceptability and​​ usefulness, recognizing its pedagogical​​​‌ value. However, they provided​ lower ratings on usability,​‌ citing the complexity of​​ metacognitive concepts and digital​​​‌ interface challenges as primary​ obstacles. The impact of​‌ these lower usability reports​​ on students' performance highlights​​​‌ critical considerations for scaling​ educational interventions. While teachers​‌ appreciated the theoretical foundations​​ and goals of the​​​‌ training, the cognitive demands​ of simultaneously managing complex​‌ pedagogical concepts and digital​​ tools during classroom implementation​​​‌ appeared to affect their​ delivery quality, which in​‌ turn influenced student outcomes.​​

8.6.3 Machine Learning for​ Adaptive Personalization in Intelligent​‌ Tutoring Systems

Participants: Pierre-Yves​​ Oudeyer [correspondant], Hélène​​​‌ Sauzeon [correspondant], Benjamin​ Clément, Didier Roy​‌, Cécile Mazon.​​

8.6.4 Machine learning​ for adaptive cognitive training​‌

Participants: Pierre-Yves Oudeyer,​​ Hélène Sauzéon [correspondant],​​​‌ Masataka Sawayama, Benjamin​ Clément, Maxime Adolphe​‌, Marion Pech,​​ Juliette Deyts.

Because​​​‌ of its cross-cutting nature​ to all cognitive activities​‌ such as learning tasks,​​ attention is a hallmark​​​‌ of good cognitive health​ throughout life and more​‌ particularly in the current​​ context of societal crisis​​​‌ of attention. Recent works​ have shown the great​‌ potential of computerized attention​​ training for an example​​​‌ of attention training, with​ efficient training transfers to​‌ other cognitive activities, and​​ this, over a wide​​​‌ spectrum of individuals (children,​ elderly, individuals with cognitive​‌ pathology such as Attention​​ Deficit and Hyperactivity Disorders).​​ Despite this promising result,​​​‌ a major hurdle is‌ challenging: the high inter-individual‌​‌ variability in responding to​​ such interventions. Some individuals​​​‌ are good responders (significant‌ improvement) to the intervention,‌​‌ others respond variably, and​​ finally some respond poorly,​​​‌ not at all, or‌ occasionally. A central limitation‌​‌ of computerized attention training​​ systems is that the​​​‌ training sequences operate in‌ a linear, non-personalized manner:‌​‌ difficulty increases in the​​ same way and along​​​‌ the same dimensions for‌ all subjects. However, different‌​‌ subjects require in principle​​ a progression at a​​​‌ different, personalized pace according‌ to the different dimensions‌​‌ that characterize attentional training​​ exercises.

To tackle the​​​‌ issue of inter-individual variability,‌ the present project proposes‌​‌ to apply some principles​​ from intelligent tutoring systems​​​‌ (ITS) to the field‌ of attention training. In‌​‌ this context, we have​​ already developed automatic curriculum​​​‌ learning algorithms such as‌ those developed in the‌​‌ KidLearn project, which allow​​ to customize the learner's​​​‌ path according to his/her‌ progress and thus optimize‌​‌ his/her learning trajectory while​​ stimulating his/her motivation by​​​‌ the progress made. ITS‌ are widely identified in‌​‌ intervention research as a​​ successful way to address​​​‌ the challenge of personalization,‌ but no studies to‌​‌ date have actually been​​ conducted for attention training.​​​‌ Thus, whether ITS, and‌ in particular personalization algorithms,‌​‌ can optimize the number​​ of respondents to an​​​‌ attention training program remains‌ an open question.

Grounded‌​‌ state-of-the-art. 

To investigate this​​ question, we first conducted​​​‌ a systematic review aiming‌ at exploring existing methods‌​‌ in computerized CT and​​ analyzing their outcomes in​​​‌ terms of learning mechanics‌ (intra-training performance) and effectiveness‌​‌ (near, far and everyday​​ life transfer effects of​​​‌ CT) 71. A‌ search up to June‌​‌ 2023 with multiple databases​​ selecting 19 computerized CT​​​‌ studies revealed that only‌ two studies emphasized the‌​‌ favorable influence of individualization​​ on CT effectiveness, while​​​‌ five underscored its capacity‌ to enhance the training‌​‌ experience by boosting motivation,​​ engagement, and offering diverse​​​‌ learning pathways. In sum,‌ despite promising results in‌​‌ this new research avenue,​​ more research is needed​​​‌ to fully understand and‌ empirically support individualized techniques‌​‌ in cognitive training.

Distribution​​ of AI techniques depending​​​‌ on type of CT‌ studied (multi or single‌​‌ domain) from Adolphe et​​ al., 2024

Complementing the study of‌​‌ adaptive methods applied to​​ cognitive training, we have​​​‌ attempted through a review‌ of the subjective literature‌​‌ to gain a better​​ understanding of the Multiple​​​‌ Object Tracking (MOT) task,‌ which seems to have‌​‌ the best results in​​ terms of attentional training​​​‌ efficiency in young and‌ older adults. Our investigation‌​‌ pursues three main objectives:​​ (1) identifying the cognitive​​​‌ processes influenced by each‌ adjustable parameter of the‌​‌ MOT task; (2) determining​​ which parameters, when progressively​​​‌ adapted during repeated MOT‌ practice, produce the greatest‌​‌ enhancements in task performance;​​ and (3) evaluating how​​​‌ improvements in MOT performance‌ translate into effective transfer‌​‌ effects, including practical, real-world​​​‌ outcomes. The evidence suggests​ that the MOT task​‌ involves a nuanced interplay​​ of visual processing, attentional​​​‌ resources, and working memory,​ shaped by the intrinsic​‌ properties of the objects​​ and the task conditions.​​​‌ The results of this​ work highlight that: (1)​‌ Multiple cognitive mechanisms are​​ identified as active in​​​‌ the task (divided and​ sustained attention; foveal and​‌ peripheric attention ; automatic​​ and controlled inhibition, etc.​​​‌ ); (2) a limited​ number of studies have​‌ actually implemented the MOT​​ task in computer-assisted cognitive​​​‌ training; and (3) tIt's​ the near (attention tasks)​‌ and far (other cognitive​​ tasks) effects that are​​​‌ well documented as positive​ outcomes of MOT-based training​‌ while there is a​​ scarcity of research that​​​‌ has thoroughly analyzed the​ ecological effects of attentional​‌ training, namely the potential​​ transfer effects in everyday​​​‌ life (paper in progress).​

Randomized and​​​‌ controlled Trial in Young​ and Olders adults :​‌ Predifined vs. ZPDES condition.​​ 

Utilizing these tools, a​​​‌ pilot study campaign was​ conducted to evaluate the​‌ impact of our AI-based​​ personalized cognitive training program.​​​‌ The first pilot experiment​ involved n=27 participants and​‌ aimed to compare the​​ effectiveness of a cognitive​​​‌ training program using a​ linear difficulty management procedure​‌ (staircase procedure) to a​​ program using an ITS​​​‌ for difficulty manipulation. The​ online training lasted for​‌ 10 hours over a​​ period of 2 weeks.​​​‌ The results indicated that​ the ITS-based intervention produced​‌ diverse learning trajectories compared​​ to the linear procedure​​​‌ 38, leading to​ broader improvements in pre-post​‌ cognitive assessment. However, no​​ significant differences were observed​​​‌ in subjective measures of​ motivation and engagement between​‌ the two groups. Subsequent​​ to this initial experiment,​​​‌ two pilot studies (n=11​ and n=10, respectively) were​‌ conducted with the goal​​ of enhancing motivation and​​​‌ engagement in the game.​ The first study implemented​‌ gamified components such as​​ scores and feedback, while​​​‌ the second study examined​ hyperparameter updates to the​‌ ITS. The analysis of​​ learning trajectories, learning outcomes,​​​‌ and subjective measures yielded​ promising results in favor​‌ of the AI-based personalized​​ procedure.

Different learning trajectories​​ for a selected participant​​​‌ in the staircase group‌ (left) and the ITS‌​‌ group (right). The color​​ of a dot indicates​​​‌ the initial presentation of‌ the parameter value, while‌​‌ the size of the​​ dot represents the frequency​​​‌ of the parameter value.‌

Building on‌ the preliminary findings, we‌​‌ expanded our research scope​​ with a more comprehensive​​​‌ experimental setup involving two‌ distinct studies. The first‌​‌ study encompassed 64 young​​ adults, sourced through the​​​‌ Prolific platform, while the‌ second study consisted of‌​‌ 50 older adults, recruited​​ from the "Université du​​​‌ temps libre". Our experimental‌ methodology mirrored that of‌​‌ our initial pilot studies,​​ with a notable enhancement:​​​‌ the integration of new‌ gamified elements (including mini-story‌​‌ creation and new visual​​ content) aimed at boosting​​​‌ participant motivation and engagement.‌

a) The MOT task.‌​‌ (b) Several visual snapshots​​ of intervention. (c) Schedule​​​‌ proposed to participants

The data analysis‌​‌ encompassed three primary dimensions:​​ initially, an exploratory phase​​​‌ to delineate learning trajectories‌ between control and intervention‌​‌ groups; subsequently, a comparative​​ analysis of pre- and​​​‌ post-test performance on the‌ cognitive battery; and lastly,‌​‌ an examination of participants'​​ self-reported experiences during training,​​​‌ providing insights into their‌ subjective perceptions of the‌​‌ experiment.

The pilot studies'​​ preliminary outcomes were corroborated​​​‌ in these larger sample‌ groups. Notably, learning trajectories‌​‌ exhibited greater diversity in​​ the group undergoing the​​​‌ intervention procedure. This group‌ also demonstrated a more‌​‌ pronounced improvement across a​​ wider range of cognitive​​​‌ assessment tasks. Although participants‌ engaging in the personalized‌​‌ cognitive training reported a​​ higher cognitive load via​​​‌ questionnaires, the levels of‌ engagement and frustration did‌​‌ not significantly differ between​​ the two groups.

The​​​‌ results showed that ZPDES‌ could be more effective‌​‌ than a control condition,​​ with improved performance on​​​‌ trained tasks in both‌ studies, underlining the benefits‌​‌ of individualized training paths.​​ However, motivation and engagement​​​‌ were lower in the‌ groups using ZPDES, probably‌​‌ due to cognitive load​​ and metacognitive factors. Overall,​​​‌ individualizing cognitive training through‌ systems like ZPDES provides‌​‌ a promising direction for​​ future research by providing​​​‌ automatic methods for taking‌ individual differences into account‌​‌ in CT programs while​​ respecting methodological standards for​​​‌ evaluating the effectiveness of‌ CT. As a result,‌​‌ our work contributes to​​ the growing body of​​​‌ knowledge in both ITS‌ and CT domains while‌​‌ stressing the crucial role​​ of challenges related to​​​‌ motivation and engagement to‌ optimize the effectiveness of‌​‌ these individualized approaches for​​ cognitive and educational outcomes.​​​‌

As part of the‌ creation of the new‌​‌ University Hospital Institute (UHI)​​ VBHI (VASCULAR BRAIN HEALTH​​​‌ INSTITUTE), we aim to‌ develop and test a‌​‌ personalized, multimodal digital therapeutic​​​‌ approach to slow down​ the functional consequences of​‌ small vessel disease. More​​ specifically:

• Evaluate the impact​​​‌ of personalized cognitive training​ compared to non-personalized conditions​‌ (comparative efficacy).
• Identify potential​​ ElectroEncephaloGraphic (EEG) biomarkers that​​​‌ reflect cognitive activity impacted​ by small vessel disease​‌ and could later (in​​ a subsequent study) be​​​‌ used as targets for​ exploratory EEG neurofeedback therapy.​‌
• Identify brain areas to​​ target for delivering non-invasive​​​‌ HD-tACS electrical stimulation, using​ previously acquired MRI data.​‌
• Evaluate the impact of​​ this stimulation on brain​​​‌ activity, neural synchronization, and​ cognitive performance.

To achieve​‌ this, 80 participants from​​ the SHIVA cohort (n=80)​​​‌ will be divided into​ two subgroups according to​‌ the severity of the​​ disease:

• Severe group: presenting​​​‌ multiple lesions on MRI​
• Non-severe group: presenting a​‌ few lesions

These groups​​ will then be further​​​‌ divided based on the​ type of training: personalized​‌ tests (ZPDES) versus standard​​ tests.

justification=centering,margin=1cm

SHIVA study​​​‌ protocol and materials.

During​​ the pre- and post-training​​​‌ sessions, participants will perform​ cognitive tests on a​‌ computer. Participants will be​​ equipped with an EEG​​​‌ headset, which, combined with​ a tACS stimulator, will​‌ allow for both brain​​ activity recording and stimulation.​​​‌

We are carrying out​ an ancillary study with​‌ Myra Fernandez's laboratory in​​ Canada, thanks to my​​​‌ participation with the Inria​ Curiositytech international associate team.​‌ We have proposed to​​ collaboratively analyze certain data​​​‌ and dimensions of interest​ in our respective laboratories​‌ (e.g. physical activities) associated​​ with the cognitive training​​​‌ proposed in the SHIVA-DTX-COG​ project.

8.6.5 ToGather : Interactive​​ website to foster collaboration​​​‌ among stakeholders of school​ inclusion for pupils with​‌ neurodevelopmental disorders

Participants: Hélène​​ Sauzéon [correspondant], Cécile​​​‌ Mazon, Eric Meyer​, Isabeau Saint-Supery,​‌ Christelle Maillart [Uni. Liège,​​ Belgium], Kamélia Belassel​​​‌, Mathieu Périé,​ Valentin Strahm.

Sustain​‌ and support the follow-up​​ of the school inclusion​​​‌ of children with neurodevelopmental​ disorders (e.g., autism, attention​‌ disorders, intellectual deficiencies) has​​ become an emergency :​​​‌ the higher is the​ school level, the lower​‌ is the amount of​​ schooled pupils with cognitive​​​‌ disabilities.

Technology-based interventions to​ improve school inclusion of​‌ children with neurodevelopmental disorders​​ have mostly been individual​​​‌ centered, focusing on their​ socio-adaptive, and cognitive impairments​‌ and implying they have​​ to adapt themselves in​​ order to fit in​​​‌ our society's expectations. Although‌ this approach centered on‌​‌ the normalization of the​​ person has some advantages​​​‌ (reduction of clinical, symptoms),‌ it carries social stereotypes‌​‌ and misconceptions of cognitive​​ disability that are not​​​‌ respectful of the cognitive‌ diversity and intrinsic motivations‌​‌ of the person, and​​ in particular of the​​​‌ student's wishes in terms‌ of school curriculum to‌​‌ achieve his or her​​ future life project 129​​​‌.

The "ToGather" project‌ aims at enlightening the‌​‌ field of educational technologies​​ for special education by​​​‌ proposing an approach centered‌ on the educational needs‌​‌ of the students and​​ bringing a concerted and​​​‌ informed answer between all‌ the stakeholders including the‌​‌ student and all their​​ support spheres (family, school,​​​‌ medico-social care). To this‌ end, ToGather project that‌​‌ emanates from participatory design​​ methods, primarily consists of​​​‌ having developed a pragmatic‌ tool (interactive website) to‌​‌ help students with cognitive​​ disability and their caregivers​​​‌ to formalize and to‌ visualize the repertoire of‌​‌ academic skills of the​​ student and to make​​​‌ it evolve according to‌ his or her proximal‌​‌ zone of development (in​​ the sense of Vygotsky)​​​‌ on the one hand,‌ and to the intrinsic‌​‌ motivations of the student​​ (his or her own​​​‌ educational and life project)‌ on the other 126‌​‌.

This project is​​ in partnership with the​​​‌ School Academy of Bordeaux‌ of the French Education‌​‌ Minestery, the ARI association,​​ the Centre of Autism​​​‌ of Aquitaine. It is‌ funded by the FIRAH‌​‌ (foundation) and the Nouvelle-Aquitaine​​ Region (see the dedicated​​​‌ webpages).

First, usability‌ studies have been conducted‌​‌ for evaluating ergonomic qualities​​ of the ToGather website,​​​‌ yielding positive resultats in‌ French and Belgian contexts.‌​‌ Then, we conducted a​​ large field-study to assess​​​‌ the effectiveness of the‌ tool in helping stakeholders‌​‌ to support children with​​ neurodevelopmental disorders (NDD)  155​​​‌  153  154.

The‌ study protocol consisted in‌​‌ a longitudinal non-randomized controlled​​ trial, with baseline, 3-months,​​​‌ and 6-months fllow-up assessments.‌ The recruitment was conducted‌​‌ across the entire French​​ territory. Our local partners​​​‌ facilitated the dissemination of‌ the call for participation‌​‌ in Gironde and provided​​ us with contacts to​​​‌ extend it to other‌ regions. Additionally, a recruitment‌​‌ campaign through social media​​ was carried out to​​​‌ communicate about the study‌ and encourage participants to‌​‌ test the ToGather tool.​​

As the tool was​​​‌ designed to support co-educational‌ process between parents and‌​‌ professionals, a support team​​ had to consist of​​​‌ at least two stakeholders,‌ including at least one‌​‌ of the parents. Initially,​​ 157 participants were recruited​​​‌ in 37 support teams,‌ but 30 individuals did‌​‌ not answer to baseline​​ questionnaire, leading to the​​​‌ exclusion of 11 support‌ teams. After baseline assessment,‌​‌ 13 support teams were​​ allocated to the experimental​​​‌ condition (ToGather app) and‌ 11 to the control‌​‌ condition (usual follow-up).

Primary​​ outcomes measures covered stakeholders’​​​‌ relationships, self-efficacy, and attitudes‌ towards inclusive education, while‌​‌ secondary outcomes measures were​​ related to stakeholders’ burden​​​‌ and quality of life,‌ as well as children’s‌​‌ school well-being and quality​​​‌ of life.

As the​ study ended recently, data​‌ analysis is still ongoing.​​ Preliminary results after 3​​​‌ months of use showed​ encouraging results with an​‌ improvement in communication between​​ stakeholders and their respective​​​‌ quality of life (paper​ in progress)

8.6.6 Curious​‌ and therefore not overloaded​​ : Study of the​​​‌ links between curiosity and​ cognitive load in learning​‌ mediated by immersive technologies​​

Participants: Hélène Sauzéon [correspondant]​​​‌, Matisse Poupard,​ André Tricot [Cosupervisor -​‌ Univ. Montpellier], Florian​​ Larrue [Industrialist - Le​​​‌ Catie].

Conducted in​ collaboration with CATIE (industrial​‌ partner) and the EPSYLON​​ laboratory at the University​​​‌ of Montpellier (under the​ supervision of Prof. André​‌ Tricot), this research program​​ was initiated in April​​​‌ 2022 and defended on​ September 11th,​‌ 2025. It pursued two​​ main objectives:

• To establish​​​‌ theoretical links between cognitive​ load theory and models​‌ of curiosity-driven learning.
• To​​ experimentally examine how the​​​‌ choice of educational technology​ modulates the relationship between​‌ pedagogical approaches (guided instruction​​ vs. exploration) and learner​​​‌ expertise.

To address these​ objectives, the thesis was​‌ structured into three main​​ phases.

Experimental Research in XR-Based​​ Anatomy Learning. 

Two experimental​​​‌ studies were conducted in​ 2023 with 131 second-year​‌ medical students and replicated​​ in 2024 with 164​​​‌ medical students from the​ second to fifth years.​‌

The first experiment investigated​​ whether supporting students’ drawing​​​‌ activity during lectures using​ augmented and mixed reality​‌ could reduce cognitive load​​ and enhance motivation. Participants​​​‌ followed a 20-minute neuroanatomy​ video lecture while simultaneously​‌ reproducing drawings demonstrated by​​ the instructor. Four experimental​​​‌ conditions were compared:

• Spatial​ Augmented Reality (SAR): A​‌ digital overlay of the​​ anatomical structure was projected​​​‌ onto paper, allowing learners​ to trace it using​‌ a projector and tracking​​ system.
• Mixed Reality (MR):​​​‌ The digital overlay was​ displayed through a HoloLens​‌ 2 headset.
• Mixed Reality​​ with 3D Model (MR+3D):​​​‌ In addition to the​ digital overlay, learners could​‌ manipulate a 3D anatomical​​ model.
• Control Condition: No​​​‌ digital overlay was provided.​

Experimental conditions for experiment​‌ 1 : Support Drawing​​ with Augmented Reality

Results from the​​​‌ 2023 dataset showed that​ both AR- and MR-supported​‌ drawing conditions significantly reduced​​ extraneous cognitive load, increased​​​‌ intrinsic motivation, and improved​ drawing accuracy. However, no​‌ significant differences in knowledge​​ acquisition were observed between​​​‌ conditions. Notably, in the​ stereoscopic 3D visualization condition,​‌ learners with higher intrinsic​​ motivation exhibited poorer learning​​​‌ outcomes, possibly due to​ increased attentional focus on​‌ system interaction rather than​​ conceptual understanding. Visuospatial ability​​ and prior knowledge moderated​​​‌ the effectiveness of AR‌ and MR interventions, with‌​‌ more experienced learners benefiting​​ the most. These results​​​‌ are reported in a‌ manuscript currently under review‌​‌ in the Journal of​​ Computing in Higher Education​​​‌65.

The second‌ experiment explored a different‌​‌ learning paradigm using virtual​​ reality (VR), manipulating levels​​​‌ of interactivity and instructional‌ guidance. This design enabled‌​‌ the examination of how​​ exploration and embodied interaction​​​‌ with a 3D anatomical‌ model affect learning outcomes,‌​‌ cognitive load, and curiosity.​​

Experimental conditions for experiment​​​‌ 2 : Embodied learning‌ in virtual reality, effect‌​‌ of interactivity

Analyses,​​ published in Computers &​​​‌ Education38, showed‌ that VR conditions led‌​‌ to superior learning performance,​​ particularly in the passive​​​‌ and active interaction conditions.‌ These conditions were associated‌​‌ with higher intrinsic motivation​​ and a more optimized​​​‌ cognitive load profile. Moreover,‌ intrinsic motivation was positively‌​‌ correlated with germane cognitive​​ load (i.e., cognitive resources​​​‌ devoted to learning) and‌ negatively correlated with extraneous‌​‌ cognitive load. In other​​ words, highly motivated learners​​​‌ experienced fewer irrelevant cognitive‌ demands, allowing them to‌​‌ allocate more resources to​​ meaningful learning processes.

Following​​​‌ the systematic review, which‌ highlighted the lack of‌​‌ reliable and context-sensitive measures​​ of intrinsic motivation in​​​‌ XR research, a third‌ study leveraged a key‌​‌ affordance of VR: the​​ continuous collection of behavioral​​​‌ data. By analyzing head‌ and hand movements during‌​‌ the neuroanatomy learning task,​​ this study aimed to​​​‌ identify implicit behavioral indicators‌ of curiosity and cognitive‌​‌ engagement. Results showed that​​ increased hand movement was​​​‌ associated with lower intrinsic‌ motivation, whereas greater head‌​‌ movement was positively associated​​ with both germane cognitive​​​‌ load and intrinsic motivation,‌ suggesting deeper cognitive engagement.‌​‌ Additionally, movement entropy emerged​​ as a significant predictor​​​‌ of curiosity-driven learning, highlighting‌ its potential as an‌​‌ implicit marker of learning-related​​ behaviors in immersive environments.​​​‌ These findings are presented‌ in a manuscript currently‌​‌ under review in the​​ International Journal of Human–Computer​​​‌ Studies37.

Illustration‌ of movement entropy calculations‌​‌ in the virtual environment​​

Toward a Unified Model:​​ Cognitive Load, Motivation, and​​​‌ Expertise. 

Building on the‌ empirical results of the‌​‌ previous studies, which revealed​​ dynamic interactions between cognitive​​​‌ load and intrinsic motivation,‌ this final phase addressed‌​‌ the second overarching objective​​ of the thesis: the​​​‌ empirical integration of cognitive‌ load theory and the‌​‌ Learning Progress Hypothesis. Using​​ structural equation modeling (SEM),​​​‌ this study tested a‌ comprehensive model describing the‌​‌ relationships between XR technologies,​​ cognitive load, intrinsic motivation,​​​‌ perceived learning progress, and‌ learning outcomes.

Results indicated‌​‌ that both AR and​​​‌ VR significantly reduced extraneous​ cognitive load and increased​‌ intrinsic motivation. However, intrinsic​​ motivation did not directly​​​‌ predict immediate learning performance.​ Instead, extraneous cognitive load​‌ negatively affected perceived learning​​ progress and autonomy, which​​​‌ in turn predicted intrinsic​ motivation, revealing a key​‌ mediating pathway.

Resulting model​​ from the SEM.

Overall,​‌ these findings demonstrate that​​ unnecessary cognitive demands not​​​‌ only hinder learning efficiency​ but also disrupt learners’​‌ perceived progress and sense​​ of control, thereby undermining​​​‌ curiosity and intrinsic motivation.​ This work contributes to​‌ a unified theoretical framework​​ by showing how optimizing​​​‌ extraneous cognitive load in​ XR environments supports both​‌ cognitive efficiency and curiosity-driven​​ learning. The results are​​​‌ presented in a manuscript​ currently under review in​‌ Educational Psychology Review66​​.

8.6.7 Self-determination-driven digital services​ for supporting aging-in place​‌ and well-being: a study​​ of relationships between longitudinal​​​‌ data from smart home​ and clinical data

Participants:​‌ Hélène Sauzéon, Juliette​​ Deyts, Lucile Dupuy​​​‌, Rafik Belloum.​

This work relies on​‌ longitudinal data collected from​​ frail older adults living​​​‌ alone at home who​ used the HomeAssist ambient​‌ assisted living platform for​​ up to 24 months.​​​‌ HomeAssist was designed according​ to a self-determination and​‌ user-centered approach, covering three​​ domains of need: daily​​​‌ activities, home safety, and​ social participation. The objective​‌ of this research is​​ to analyze relationships between​​​‌ clinical data (e.g., cognitive​ assessments, frailty, autonomy, self-determination)​‌ and usage-related data (user​​ experience questionnaires, usage diaries,​​​‌ and actimetric data derived​ from environmental sensors), in​‌ order to both assess​​ the benefits of assistive​​​‌ and monitoring services and​ explore the predictive value​‌ of sensor-based data for​​ explaining clinical outcomes.

A​​​‌ first study focused on​ identifying factors influencing user​‌ experience (UX) and long-term​​ adoption of HomeAssist, based​​​‌ on data from 131​ participants. Despite a user-centered​‌ design, long-term adoption remained​​ limited, with only 18​​​‌ users continuing after 24​ months and 38 requesting​‌ removal within the first​​ six months. Regression analyses​​​‌ showed that UX dimensions​ were mainly predicted by​‌ other UX dimensions rather​​ than by individual health​​​‌ or psychosocial characteristics. In​ contrast, long-term adoption was​‌ weakly predicted by level​​ of education and computer​​​‌ ownership, suggesting that while​ user-centered design may reduce​‌ the impact of individual​​ characteristics on user experience,​​​‌ adoption remains influenced by​ digital literacy and social​‌ inequalities.

Overall, these activities​​ contribute to the design​​ of a user-centered visualization​​​‌ tool intended for clinicians‌ (psychologists and physicians), enabling‌​‌ them to better understand​​ the links between long-term​​​‌ usage data and clinical‌ evolution, and to detect‌​‌ early “weak signals” of​​ decline (e.g., changes in​​​‌ sleep patterns), thereby facilitating‌ timely and targeted interventions.‌​‌

8.7.1​​​‌ Design of an Interactive‌ Software for Automated Discovery‌​‌ in Complex Systems

Participants:​​ Clément Romac [correspondant],​​​‌ Zacharie Bugaud, Clément‌ Moulin-Frier, Pierre-Yves Oudeyer‌​‌.

We further developed​​ our Automated Discovery software​​​‌ and particularly focused on‌ adding and experimenting with‌​‌ new systems.

Our public​​ software now features more​​​‌ than ten examples ranging‌ from artificial life, to‌​‌ physics or protein docking.​​ The software was publicly​​​‌ released in 2024: presentation‌ thread.

Technical architecture‌​‌ of our software.

8.7.2‌ Discovering Sensorimotor Agency in‌​‌ Cellular Automata using Diversity​​ Search

Participants: Gautier Hamon​​​‌ [correspondant], Mayalen Etcheverry‌, Bert Chan,‌​‌ Clément Moulin-Frier, Pierre-Yves​​ Oudeyer.

As a​​​‌ continuation of the previous‌ projects in Automated Discovery‌​‌ in Self-Organizing Systems, we​​ have been working on​​​‌ expanding the set of‌ discoveries of possible structures‌​‌ in continuous CAs such​​ as Lenia 82,​​​‌ 81, and in‌ particular we have been‌​‌ interested to search for​​ emerging agents with sensorimotor​​​‌ capabilities. Understanding what has‌ led to the emergence‌​‌ of life and sensorimotor​​ agency as we observe​​​‌ in living organisms is‌ a fundamental question. In‌​‌ our work, we initially​​ only assume environments made​​​‌ of low-level elements of‌ matter (called atoms, molecules‌​‌ or cells) locally interacting​​ via physics-like rules. There​​​‌ is no predefined notion‌ of agent embodiment and‌​‌ yet we aim to​​ answer the following scientific​​​‌ question: is it possible‌ to find environments in‌​‌ which there exists/emerge a​​ subpart that could be​​​‌ called a sensorimotor agent‌?

We use Lenia‌​‌ continuous cellular automaton as​​ our artificial "world" 81​​​‌. We introduce a‌ novel method based on‌​‌ gradient descent and curriculum​​ learning combined within an​​​‌ intrinsically-motivated goal exploration process‌ (IMGEP) to automatically search‌​‌ parameters of the CA​​ rule that can self-organize​​​‌ spatially localized 1 and‌ moving patterns 2 within‌​‌ Lenia. The IMGEP defines​​ an outer exploratory loop​​​‌ (generation of training goal/loss)‌ and an inner optimization‌​‌ loop (goal-conditioned). We use​​ a population-based version of​​​‌ IMGEP 12,91‌ but introduce two novel‌​‌ elements compared to previous​​ papers in the IMGEP​​​‌ literature. First, whereas previous‌ work in 29 and‌​‌ 10 used a very​​ basic nearest-neighbor goal-achievement strategy,​​​‌ our work relies on‌ gradient descent for the‌​‌ local optimization of the​​ (sensitive) parameters of the​​​‌ complex system, which has‌ shown to be very‌​‌ powerful. To do so​​ we made a differentiable​​​‌ version of the Lenia‌ framework, which is also‌​‌ a contribution of this​​ work. Secondly, we propose​​​‌ to control subparts of‌ the environmental dynamics with‌​‌ functional constraints (through predefined​​ channels and kernels in​​​‌ Lenia) to build a‌ curriculum of tasks; and‌​‌ to integrate this stochasticity​​​‌ in the inner optimization​ loop. This has shown​‌ central to train the​​ system to emerge sensorimotor​​​‌ agents that are robust​ to stochastic perturbations in​‌ the environment. In particular,​​ we focus on modeling​​​‌ obstacles in the environment​ physics and propose to​‌ probe the agent sensorimotor​​ capability as its performance​​​‌ to move forward under​ a variety of obstacle​‌ configurations. We also provide​​ in this work tests​​​‌ and metrics to measure​ the robustness of the​‌ obtained agents.

Robustness test​​ to harder/unseen obstacle configurations:​​​‌ straight wall, bigger obstacle,​ dead ends.

Change of scale​‌ changing the kernel size​​ and initialization, the grid​​​‌ is the same size​ in both

While many​​ complex behaviors have already​​​‌ been observed in Lenia,​ among which some could​‌ qualify as sensorimotor behaviors,​​ they have so far​​​‌ been discovered "by chance"​ as the result of​‌ time-consuming manual search or​​ with simple evolutionary algorithms.​​​‌ Our method provides a​ more systematic way to​‌ automatically learn the CA​​ rules leading to the​​​‌ emergence of basic sensorimotor​ structures, as shown in​‌ Figure 48. Moreover,​​ we investigated and provided​​​‌ ways to measure the​ (zero-shot) generalization of the​‌ discovered sensorimotor agents to​​ several out-of-distribution perturbations that​​​‌ were not encountered during​ training. Impressively, even though​‌ the agents still fail​​ to preserve their integrity​​​‌ in certain configurations, they​ show very strong robustness​‌ to most of the​​ tested variations. The agents​​​‌ are able to navigate​ in unseen and harder​‌ environmental configurations while self-maintaining​​ their individuality (Figure 46​​​‌). Not only the​ agents are able to​‌ recover their individuality when​​ subjected to external perturbations​​​‌ but also when subjected​ to internal perturbations: they​‌ resist variations of the​​ morphogenetic processes such that​​​‌ less frequent cell updates,​ quite drastic changes of​‌ scales as well as​​ changes of initialization (Figure​​​‌ 47). Furthermore, when​ tested in a multi-entity​‌ initialization and despite hav,ing​​ been trained alone, not​​​‌ only the agents are​ able to preserve their​‌ individuality but they show​​ forms of coordinated interactions​​​‌ (attractiveness and reproduction). Our​ results sug,gest that, contrary​‌ to the (still predominant)​​ mechanistic view on embodiment,​​​‌ biologically-inspired embodiment could pave​ the way toward agents​‌ with strong coherence and​​ generalization to out-of-distribution changes,​​​‌ mimicking the remarkable robustness​ of living systems to​‌ maintain specific functions despite​​ environmental and body perturbations​​​‌ 116. Searching for​ rules at the cell-level​‌ in order to give​​ rise to higher-level cognitive​​​‌ processes at the level​ of the organism and​‌ at the level of​​ the group of organisms​​​‌ opens many exciting opportunities​ to the development of​‌ embodied approaches in AI​​ in general.

Scatter plot​​​‌ of the agents as​ their measured performances of​‌ robustness to obstacles (y​​ axis) and speed in​​​‌ obstacles (x axis) obtained​ by IMGEP (red), random​‌ search with the same​​ compute resources as IMGEP(blue)​​ and the one from​​​‌ the original lenia paper‌ (green)

The‌ work has been released‌​‌ in 2022 as a​​ distill-like article which is​​​‌ currently hosted at this‌ link. This article‌​‌ contains an interactive demo​​ in webGL and javascript,​​​‌ as well as many‌ videos and animations of‌​‌ the results. A colab​​ notebook with the source​​​‌ code of the work‌ is publicly available at‌​‌.

In 2024, additional​​ quantitative experiments were conducted​​​‌ as well as ablations.‌ This work was published‌​‌ in 2025in the​​ Science Advances journal.

8.7.3​​​‌ Semantic Open-Endedness in Flow-Lenia‌ using Vision Language Models‌​‌ and IMGEP

Participants: Sina​​ Khajehabdollahi [correspondent], Gautier​​​‌ Hamon, Marko Cvjetko‌, Cédric Colas,‌​‌ Pierre-Yves Oudeyer, Clément​​ Moulin-Frier.

Discovering diverse​​​‌ visual patterns in continuous‌ cellular automata (CA) is‌​‌ challenging due to the​​ vastness and redundancy of​​​‌ high-dimensional behavioral spaces. Traditional‌ exploration methods like Novelty‌​‌ Search (NS) expand locally​​ by mutating known novel​​​‌ solutions but often plateau‌ when local novelty is‌​‌ exhausted, failing to reach​​ distant, unexplored regions. We​​​‌ introduce Expedition & Expansion‌ (E&E), a hybrid strategy‌​‌ where exploration alternates between​​ local novelty-driven expansions and​​​‌ goal-directed expeditions. During expeditions,‌ E&E leverages a Vision-Language‌​‌ Model (VLM) to generate​​ linguistic goals—descriptions of interesting​​​‌ but hypothetical patterns that‌ drive exploration toward uncharted‌​‌ regions. By operating in​​ semantic spaces that align​​​‌ with human perception, E&E‌ both evaluates novelty and‌​‌ generates goals in conceptually​​ meaningful ways, enhancing the​​​‌ interpretability and relevance of‌ discovered behaviors. Tested on‌​‌ Flow Lenia, a continuous​​ CA known for its​​​‌ rich, emergent behaviors, E&E‌ consistently uncovers more diverse‌​‌ solutions than existing exploration​​ methods. A genealogical analysis​​​‌ further reveals that solutions‌ originating from expeditions disproportionately‌​‌ influence long-term exploration, unlocking​​ new behavioral niches that​​​‌ serve as stepping stones‌ for subsequent search. These‌​‌ findings highlight E&E's capacity​​ to break through local​​​‌ novelty boundaries and explore‌ behavioral landscapes in human-aligned,‌​‌ interpretable ways, offering a​​ promising template for open-ended​​​‌ exploration in artificial life‌ and beyond. The project‌​‌ was published at the​​ Artificial Life 2025 conference​​​‌ 48. A summary‌ and the result visualization‌​‌ are available on the​​ project website.

8.7.4​​​‌ Exploring Flow-Lenia Universes with‌ a Curiosity-driven AI Scientist:‌​‌ Discovering Diverse Ecosystem Dynamics​​

Participants: Thomas Michel [correspondent]​​​‌, Marko Cvjetko [correspondent]‌, Gautier Hamon,‌​‌ Pierre-Yves Oudeyer, Clément​​ Moulin-Frier.

We present​​​‌ a method for the‌ automated discovery of system-level‌​‌ dynamics in Flow-Lenia—a continuous​​ cellular automaton with mass​​​‌ conservation and parameter localization—using‌ a curiosity-driven AI scientist.‌​‌ This method aims to​​ uncover processes leading to​​​‌ self-organization of evolutionary and‌ ecosystemic dynamics in CAs.‌​‌ We build on previous​​ work which uses diversity​​​‌ search algorithms in Lenia‌ to find self- organized‌​‌ individual patterns, and extend​​​‌ it to large environments​ that support distinct interacting​‌ patterns. We adapt Intrinsically​​ Motivated Goal Exploration Processes​​​‌ (IMGEPs) to drive exploration​ of diverse Flow-Lenia environments​‌ using simulation-wide metrics, such​​ as evolutionary activity, compression-based​​​‌ complexity, and multi-scale entropy.​ We test our method​‌ in two experiments, showcasing​​ its ability to illuminate​​​‌ significantly more diverse dynamics​ compared to random search.​‌ We show qualitative results​​ illustrating how ecosystemic simulations​​​‌ enable self-organization of complex​ collective behaviors not captured​‌ by previous individual pattern​​ search and analysis. We​​​‌ complement automated discovery with​ an interactive exploration tool,​‌ creating an effective human-AI​​ collaborative workflow for scientific​​​‌ investigation. Though demonstrated specifically​ with Flow-Lenia, this methodology​‌ provides a framework potentially​​ applicable to other parameterizable​​​‌ complex systems where understanding​ emergent collective properties is​‌ of interest.

This work​​ was published at the​​​‌ Artificial Life 2025 conference​ 51, with a​‌ companion website containing videos​​ of the discoveries, the​​​‌ interactive exploration tool and​ source code.

A showcase​‌ of discovered diversity while​​ searching for ecosystemic dynamics.​​​‌

8.7.5 Discovering and​​​‌ Controlling Diverse Self-Organized Patterns​ in Cellular Automata Using​‌ Autotelic Reinforcement Learning

Participants:​​ Marko Cvjetko [correspondent],​​​‌ Gautier Hamon, Pierre-Yves​ Oudeyer, Clément Moulin-Frier​‌.

Autotelic AI algorithms,​​ which pursue self-generated goals,​​​‌ have proven to be​ effective as automated discovery​‌ assistants in cellular automata.​​ Previous work in this​​​‌ domain focused on algorithms​ which produce diverse behaviors​‌ by setting the automaton’s​​ initial conditions. Here, we​​​‌ extend these methods beyond​ initial-condition search and adapt​‌ them to systems that​​ support sequences of closed-loop​​​‌ interventions. Using Lenia (a​ continuous cellular automaton) as​‌ a test environment, we​​ train goal-conditioned reinforcement learning​​​‌ agents to perform targeted​ interventions during the system’s​‌ evolution, guiding it towards​​ desired states. The resulting​​​‌ agent behaviors are robust​ and diverse, demonstrating the​‌ potential of closed-loop interaction​​ for discovery and control.​​​‌ Furthermore, we show that​ goal-conditioned RL agents performing​‌ interventions can discover novel​​ self-organising patterns and generalize​​​‌ to previously unseen and​ noisy environments. The project​‌ was presented as a​​ late-breaking abstract at the​​​‌ Artificial Life 2025 conference​ 45, and disseminated​‌ through a website.​​

CLEMENCE​ Cohort (Fondation de France​‌ and Théa Pharma)

Participants:​​ Hélène Sauzéon [correspondant],​​​‌ Cécile Mazon, Cécile​ Delcourt.

The project​‌ "Cohorte LongitudinalE sur la​​ Myopie et le développement​​​‌ oculaire dans l’ENfanCE(CLEMENCE) is​ led by C. Delcourt​‌ from the lab of​​ Bordeaux Populational Health (2M€).​​​‌ Hélène Sauzéon and Cécile​ Mazon participate to the​‌ research program with the​​ study of developemental changes​​ due to Myopa in​​​‌ visual attention.

10.1.1 Inria associate​​ team not involved in​​​‌ an IIL or an‌ international program

Participants: Helene‌​‌ Sauzéon, Edith Law​​.

10.2.1 Horizon‌​‌ Europe

Participants: Cédric Colas​​.

10.2.2 Other european programs/initiatives​​

Participants: Helene Sauzéon,​​​‌ Pierre-Yves Oudeyer, Mathias​ Grüber.

GAIMHE project​​​‌ (BPI France 2030):

GAIMHE​ is a strategic research​‌ and innovation project funded​​ by Bpifrance, coordinated by​​​‌ EvidenceB in partnership with​ the Flowers AI &​‌ Cognitive Science Laboratory at​​ Inria, Café pédagogique, and​​​‌ Association Class'Code. The project​ aims to develop next-generation​‌ intelligent tutoring systems that​​ combine the pedagogical rigor​​​‌ of traditional adaptive learning​ algorithms with the flexibility​‌ of generative AI. Current​​ educational AI technologies present​​​‌ a fundamental trade-off. Intelligent​ Tutoring Systems (ITS) offer​‌ pedagogically grounded, personalized curricula​​ through algorithms such as​​​‌ ZPDES, but require substantial​ manual content development. Conversely,​‌ generative AI provides interactional​​ flexibility yet lacks pedagogical​​​‌ structure, cannot support long-term​ curriculum personalization, and presents​‌ significant computational costs. GAIMHE​​ proposes a hybrid methodology​​​‌ structured around three axes:​ automated content generation leveraging​‌ generative AI for the​​ rapid creation of pedagogically​​​‌ validated exercises to populate​ ITS knowledge graphs; targeted​‌ generative assistance deploying optimally-sized​​ models to provide pedagogically​​ principled guidance at key​​​‌ learning moments; and advanced‌ personalization through compact student‌​‌ models capable of predicting​​ and adapting learning trajectories​​​‌ across extensive exercise spaces,‌ building on prior MAGELLAN‌​‌ research. The project benefits​​ from EvidenceB's existing infrastructure,​​​‌ which serves tens of‌ thousands of classrooms across‌​‌ primary and secondary education​​ in France. This enables​​​‌ large-scale evaluation with authentic‌ learning data. In partnership‌​‌ with Région Île-de-France, the​​ consortium will release annotated​​​‌ datasets, learning analytics tools,‌ and software components as‌​‌ open-source digital commons to​​ support France's educational technology​​​‌ ecosystem.

ANR Chaire Individuelle‌ Deep Curiosity

- PY‌​‌ Oudeyer continued to work​​ on the research program​​​‌ of this Chaire, funding‌ 2 PhDs and 3‌​‌ postdocs for five years​​ (until 2025).

ANR JCJC​​​‌ ECOCURL

- C. Moulin-Frier‌ obtained an ANR JCJC‌​‌ grant. The project is​​ entitled "ECOCURL: Emergent communication​​​‌ through curiosity-driven multi-agent reinforcement‌ learning". The project starts‌​‌ in Feb 2021 for​​ a duration of 48​​​‌ months. It will fund‌ a PhD student (36‌​‌ months) and a Research​​ Engineer (18 months) as​​​‌ well as 4 Master‌ internships (one per year).‌​‌

Projet AIxIA: "Analyse d’Interférences​​ par Intelligence Artificielle".

Pierre-Yves​​​‌ Oudeyer and Clément Moulin-Frier‌ obtained a grant from‌​‌ the call for project​​ AIRSTRIP "L'intelligence Artificielle au​​​‌ service de l'IngénieRie des‌ SysTèmes aéRonautIques et sPatiaux",‌​‌ in collaboration with the​​ IRT Saint Exupery. The​​​‌ project was accepted in‌ 2023 and will fund‌​‌ 18 months of a​​ research engineer position starting​​​‌ in 2024.

Inria Exploratory‌ Action AIDE

- Didier‌​‌ Roy is collaborator of​​ the Inria Exploratory Action​​​‌ AIDE "Artificial Intelligence Devoted‌ to Education", ported by‌​‌ Frédéric Alexandre (Inria Mnemosyne​​ Project-Team), Margarida Romero (LINE​​​‌ Lab) and Thierry Viéville‌ (Inria Mnemosyne Project-Team, LINE‌​‌ Lab). The aim of​​ this Exploratory Action consists​​​‌ to explore to what‌ extent approaches or methods‌​‌ from cognitive neuroscience, linked​​ to machine learning and​​​‌ knowledge representation, could help‌ to better formalize human‌​‌ learning as studied in​​ educational sciences. AIDE is​​​‌ a four year project‌ started middle 2020 until‌​‌ 2024 see.

Inria​​ Exploratory Action I'AM

-​​​‌ Hélène Sauzéon is co-PI‌ with P. Dragicevic of‌​‌ the Inria Exploratory Action​​ I'AM "Impact of Augmented​​​‌ Reality on Autobiographical Memory:‌ Examining Involuntary Memories and‌​‌ False Memories" (174,5k€). Starting​​ in last september, the​​​‌ aim of this Exploratory‌ Action consists to explore‌​‌ to what extent augmented​​ reality based devices can​​​‌ produce erroneous autobiographical memories,‌ and more particularly in‌​‌ vulnerable people (Children and​​ older adults or yound​​​‌ adults with low memory‌ abilities of source monitoring).‌​‌

New collaboration with Maxime​​ Derex from IAST Toulouse​​​‌

for the co-direction of‌ the PhD thesis of‌​‌ Jeremy Perez with Clément​​ Moulin-Frier and Pierre-Yves Oudeyer​​​‌ on "Interactions between intrinsically‌ motivated goal-exploration processes and‌​‌ cummulative cultural evolution" (see​​ section 8.2.2).

France​​​‌ 2030 - PPR AUTONOMIE‌ : Vieillissement Et Situations‌​‌ De Handicap - Projet​​ INNOVCare (Lechevalier S., 3,5M€)​​​‌ (2023-26)

- Hélène Sauzéon‌ and AS Rigaud will‌​‌ supervize the WP5 dedicated​​ to two care-led innovation​​​‌ experiments with assistive technologies‌ (400k € for Bordeaux).‌​‌ - Hélène Sauzéon is​​​‌ responsible of the WP3​ « Digital technology for​‌ aging in place »​​ (470k€/3,5M€), Défi 4 -​​​‌ Numérique, Innovcare (PPR Autonomie​ PIA2030, 2023-28).

VBHI project(Vascular​‌ Brain Health Institute -IHU,​​ led by S. Debette,​​​‌ 5M€)) (2023-26)

- Hélène​ Sauzéon will supervize the​‌ WP4.3 dedicated to "Explore​​ Digital Therapeutics To Slow​​​‌ Down Cognitive Decline In​ Covert Csvd" (150k€)

11.1.1 Scientific events:​​​‌ organisation

PY Oudeyer continued​ to be a member​‌ of the organization committe​​ of the Life, Structure​​​‌ and Cognition symposium series​ at IHES, France. H​‌ sauzeon continued to be​​ a member of the​​​‌ Technical Program committee of​ ACHI conference.

11.1.2 Reviewer​‌ - reviewing activities

Matisse​​ Poupard has reviewed for​​​‌ Computers & Education, Education​ and Information Technologies, Frontiers​‌ in Psychology and Frontiers​​ in Virtual Reality. Jeremy​​​‌ Perez has reviewed for​ the Judgment and Decision​‌ Making Journal and Topics​​ in Cognitive Science. Hélène​​​‌ Sauzéon has reviewed for​ ACM-CHI25 , British Journal​‌ of Psychology, Computer in​​ human behavior, and for​​​‌ Journal of Research in​ Science Teaching. Cécile Mazon​‌ has reviewed for Nature​​ Scientific Reports, Education and​​​‌ Information Technology, and BMC​ Psychology.

PY Oudeyer was​‌ a reviewer for the​​ journals Developmental Science and​​​‌ Child Development.

11.1.3 Invited​ talks

Matisse Poupard gave​‌ 3 invited talks:

• (December​​ 2025) "Technologies immersives pour​​​‌ l’enseignement : étude des​ relations entre charge cognitive​‌ et curiosité des apprenants",​​ ReSCi “Ma Recherche j’en​​​‌ parle” - Education, outils​ numériques et intelligence artificielle"​‌, ANRT, Paris
• (November​​ 2025) Round-table discussion to​​​‌ mark the end of​ the Dem’UP project,​‌ University of Poitiers
• (October​​ 2025) "Curieux et cognitivement​​​‌ engagé", Séminaire « Numérique​ pour l’Education »,​‌ R3NumED

Clément Romac gave​​ 1 invited talks:

• (May​​​‌ 2025) Invited talk at​ the ISIR lab from​‌ Sorbonne University on “Grounding​​ LLMs through curiosity-driven online​​​‌ RL”.
• (Spetember 2025) Invited​ talk at the SMILES​‌ workshop at ICDL on​​ “Grounding LLMs through curiosity-driven​​​‌ online RL”.

Marie-Sarah Desvaux​ gave 1 invited talk:​‌

• (November 2025) "Curiosity-driven learning​​ as a ZPD window​​​‌ for self-regulated learning" during​ International Society of Cultural-historical​‌ Activity Research, Southern Europe​​ and Middle Eastern Conference​​​‌ 2025, Barcelona

Hélène​ Sauzéon gave 3 invited​‌ talk:

• (October 2025) "Supporting​​ digital accessibility of MOOC​​​‌ based-learning for individuals with​ cognitive impairments: The Aïana​‌ project" Intersections: Translation, Accessibility,​​ Inclusion, Forum des Savoirs,​​​‌ MSH Dijon
• (Mai 2025)​ Les interventions de santé​‌ pour le bien-viellir des​​ personnes âgés à l'aide​​​‌ de technologies numériques Journée​ IFRATH - Troubles sociocognitifs​‌ et technologies : Perspectives​​ sur l'enfance et le​​​‌ vieillissement, 'Institut National de​ Jeunes Sourds de Paris,​‌ 254 Rue Saint-Jacques, 75005​​ Paris.
• (November 2025) The​​​‌ intrinsic motivations as design​ principles of technologies for​‌ cognition : Examples about​​ Educational Technologies and Technologies​​​‌ for aging in place,​ Institut de Psychologie, Université​‌ Paris Cité, Paris

Loris​​ Gaven gave 2 invited​​​‌ talks:

• (January 2026) "Toward​ Artificial Curiosity" at University​‌ of Padua (Online)
• (August​​ 2025) "MAGELLAN: Metacognitive predictions​​​‌ of learning progress guide​ autotelic LLM agents in​‌ large goal spaces" at​​ the Metacognitive Satellite of​​ CCN in Amsterdam.

PY​​​‌ Oudeyer gave these invited‌ talks:

• (Jan 2025) Curiosity-driven‌​‌ learning in humans: learning​​ progress, autotelic exploration and​​​‌ open-ended development, Keynote lecture‌ at the Budapest Conference‌​‌ on Cognitive Development,​​ see video.
• (Jan​​​‌ 2025) IA générative et‌ éducation: enjeux sociétaux,‌​‌ for the "Conférence jumelle​​ du Cnesco" co-organized by​​​‌ Cnesco, Canope and CARDIE‌ Charentes et Charentes-Maritime.
• (Jan‌​‌ 2025) Les enjeux sociétaux​​ de l'IA dans l'éducation​​​‌, at ETAPP-IA conference‌ on AI and education‌​‌ organized by Nouvelle Aquitaine​​ academy.
• (May 2025) IA​​​‌ générative, société et éducation:‌ les enjeux de la‌​‌ formation des futures citoyens​​, for the conference​​​‌ "Printemps de la recherche‌ en éducation", organized by‌​‌ INSPE, Paris.
• (Oct 2025)​​ How curiosity drives human​​​‌ learning, and how this‌ can be leveraged in‌​‌ educational technologies, LEAD​​ symposium organized by University​​​‌ of Tuebingen.
• (Oct 2025)‌ IA générative: enjeux, interventions‌​‌ et résultats expérimentaux,​​ for the AI seminar​​​‌ organized by DRANE from‌ Grenoble academy.
• (Dec 2025)‌​‌ with Julien Pourcel and​​ Cédric Colas, SOAR, a​​​‌ self-improving LLM-based evolutionary algorithm‌, for the ARC-Prize.‌​‌

11.1.4 Scientific expertise

Hélène​​ Sauzéon was:

• Vice-president of​​​‌ Pluridisciplinary committee (Digital sciences‌ and Humanities) of French‌​‌ National agency for Research​​ (CES 38 -Interface ANR),​​​‌ since 2025
• Member of‌ Research Council of Finland‌​‌ POC (12 projects on​​ applied computer sciences) in​​​‌ 2025
• Member of the‌ Scientific Committee of Calyxis,‌​‌ a center focused on​​ research and development of​​​‌ technological solutions to prevent‌ daily accidents through public‌​‌ laboratory-enterprise collaborations, since 2019.​​
• Expert for grant applications:​​​‌ Evaluation of 2CIFRE-ANRT PhD‌ proposals ; Evaluation of‌​‌ 1 GATES (Grenoble ATtractiveness​​ and ExcellenceS) proposal for​​​‌ the SHS Cluster of‌ Université Grenoble in 2025.‌​‌
• Member of committe for​​ a permanent Professor position​​​‌ in psychology at the‌ university of Bordeaux
• Member‌​‌ of committe for a​​ permanent Assistant Professor position​​​‌ in Occupational Science (91‌ section) at the university‌​‌ of Limoges

PY Oudeyer​​ was:

• a reviewer and​​​‌ expert for ANR (National‌ Research Agency) as well‌​‌ as for the European​​ Research Council (ERC), the​​​‌ Cyprus Research Council and‌ for the Swedish Foundation‌​‌ for Strategic Research.
• an​​ invited expert for the​​​‌ "Curiosity Convening" event organized‌ by the Scratch Foundation‌​‌ at OECD, Paris.
• invited​​ to be a member​​​‌ of the scientific council‌ of the "Main à‌​‌ la Pate" foundation.​​
• a member of the​​​‌ GT "IA et éducation"‌ at Conseil Scientifique de‌​‌ l'Education Nationale.

Cécile Mazon​​ reviewed one proposal for​​​‌ ANR-AAPG JCJC.

11.1.5 Research‌ administration

Hélène Sauzéon was:‌​‌

• Member of the Research​​ Committee of IMT Atlantique​​​‌ since 2025, working to‌ promote Human and Social‌​‌ Sciences (SHS) in engineering​​ education.
• Co-organizer (Inria) since​​​‌ 2024 of the annual‌ "JS & GT 'Handicap'"‌​‌ (Thematic Days & Working​​ Groups on Disability) and​​​‌ contributor to the consultation‌ for Inria's 2025 Disability‌​‌ Roadmap.
• Member of the​​ extented "BCP" of BSO​​​‌ Inria centre, since 2020.‌ Advisory roles for the‌​‌ center's “surrounding” scientific policy​​ and strategy, recruitment of​​​‌ permanent researchers, especially ,‌ monitoring and assistance in‌​‌ setting up Inria teams,​​​‌ organization of intern scientific​ events, writting support to​‌ communication staff for popularization​​ contents on AI, disabilty,​​​‌ health and Education, etc.​
• PIQ Referal for the​‌ centre Inria of Univ.​​ of Bordeaux covering 3​​​‌ universities (Bordeaux, LaRochelle, Limoges),​ since 2024. My role​‌ is twofold: 1) to​​ follow up and help​​​‌ site referents and applicants​ to define and draft​‌ projects, while ensuring compliance​​ with PIQ program policy,​​​‌ i.e. close dialogue with​ PIQ staff, and 2)​‌ to inform the center's​​ scientific management of applications​​​‌ in progress in New​ Aquitaine via a dedicated​‌ "pad", and their positioning​​ in relation to the​​​‌ PIQ program's national results.​
• Referal of Education topic​‌ for the centre Inria​​ of Univ. of Bordeaux​​​‌ (covering 3 teams :​ Bivouac, Flowers, Mnemosyne), and​‌ for which I'm the​​ centre proxis at RTP​​​‌ CNRS Éducation.
• Head of​ an Associate Inria Team-​‌ CuriousTech Inria-UW–Univ. Waterloo (Canada),​​ since 2023. The multi-disciplinary​​​‌ program (Prof. M. Fernandes'​ psychology lab, and Edith​‌ Law's HCI lab) involves​​ designing innovative interactive systems​​​‌ for education and cognitive​ health at all ages,​‌ with the singularity of​​ leveraging intrinsic motivations (self-determination​​​‌ and curiosity) as reinforcers​ of human performance.
• Member​‌ of the Direction Committee​​ of IFR Handicap (Inserm)​​​‌ labelled Fedhra since 2023,​ since 2019
• Member of​‌ the Direction Committee of​​ BIND - centre excellence​​​‌ BIND de Bordeaux, since​ 2019
• Member of the​‌ scientific Committee of SOUND​​ - centre excellence TND​​​‌ Bordeaux, since 2025
• Resp.​ of Research Axis on​‌ Innovating Interventions at ACTIVE​​ Team (BPH Lab), since​​​‌ 2022

Cécile Mazon is:​

• Co-responsible of the Digital​‌ Tools workpackage of the​​ PIA Atypie-Friendly
• Local contact​​​‌ for Inria HandiTechLab
• Member​ of the Digital Tools​‌ axis of the Bordeaux​​ Excellence center for Neurodevelopmental​​​‌ disorders (SOUND project)

PY​ Oudeyer was:

• head of​‌ the Flowers AI &​​ CogSci lab
• member of​​​‌ the piloting committee of​ the France 2030 BPI​‌ project GAIMHE
• representative of​​ Inria in the piloting​​​‌ committee of the Nouvelle​ Aquitaine Research Network on​‌ Educational Technologies (R3NumEd)

11.2.1 Teaching​‌

Cécile Mazon is responsible​​ of:

• Cognitive science curriculum​​​‌ in MIASHS bachelor (Mathematics​ and Computer Science applied​‌ to Social and Human​​ Sciences) - since 2024​​​‌
• Technology, Ergonomy, Cognition, Disability​ curriculum in Cognitive Sciences​‌ master, since 2022
• Apprenticeship​​ academic coordination for Technology,​​​‌ Ergonomy, Cognition, Disability curriculum​ in Cognitive Sciences master​‌ - since 2023

Leslie​​ Tricoche , as ATER,​​​‌ gave the following courses:​

• L2 MIASHS - UFR​‌ Sciences and Technology, Bordeaux​​ University: Neurobiology (lectures and​​​‌ tutorials)
• L3 MIASHS -​ UFR Sciences and Technology,​‌ Bordeaux University: Neuropathology (lectures​​ and tutorials)
• M1 Cognitive​​​‌ Sciences - UFR Sciences​ and Technology, Bordeaux University:​‌ Cognitive functions in situations​​ and disabilities (lectures and​​​‌ tutorials)
• M2 Cognitive Sciences​ - UFR Sciences and​‌ Technology, Bordeaux University: Multiple​​ forms of the profession​​​‌ (lectures and tutorials)

Marie-Sarah​ Desvaux , as Teaching​‌ Assistant, gave the following​​ courses:

• M2 Cognitive Sciences​​​‌ - UFR Sciences and​ Technology, Bordeaux University: Multiple​‌ forms of the profession​​ - Project management
• L3​​ MIASHS - UFR Sciences​​​‌ and Technology, Bordeaux University:‌ Web Accessibility

Juliette Deyts‌​‌ , as Teaching Assistant,​​ gave the following courses:​​​‌

• M2 Cognitive Sciences -‌ UFR Sciences and Technology,‌​‌ Bordeaux University: Disability, Autonomy,​​ Cognition and Technology

Matisse​​​‌ Poupard , as ATER,‌ gave the following courses:‌​‌

• L1 MIASHS - UFR​​ Sciences and Technology, Bordeaux​​​‌ University: Introduction to Cognitive‌ Science
• L2 MIASHS -‌​‌ UFR Sciences and Technology,​​ Bordeaux University: Neurological Foundations,​​​‌ Cognitive Fundamentals, and Learning‌
• L3 MIASHS - UFR‌​‌ Sciences and Technology, Bordeaux​​ University: Knowledge and Representations,​​​‌ Language, and Natural Language‌ Processing
• M1 Cognitive Sciences‌​‌ - UFR Sciences and​​ Technology, Bordeaux University:
  • Scientific​​​‌ Foundations
  • Cognitive Functions in‌ Situations and Disabilities
• M2‌​‌ Cognitive Sciences - UFR​​ Sciences and Technology, Bordeaux​​​‌ University:
  • Disability, Activity, Cognition,‌ Technology
  • Multiple Forms of‌​‌ the Profession
  • Virtual Reality,​​ Interaction, and Health Applications​​​‌

Cécile Mazon , as‌ assistant professor, gave lectures‌​‌ and tutorials (280hETD) in​​ cognitive sciences to students​​​‌ in MIASHS bachelor (L1-2-3)‌ and Cognitive sciences master‌​‌ (M1-M2). Key teaching topics​​ include introduction to cognitive​​​‌ sciences, cognitive psychology (main‌ cognitive functions, experimental methods),‌​‌ cognitive sciences applied to​​ disability and/or technology design,​​​‌ as well as methodology‌ and statistics.

Helene Sauzéon‌​‌ participated to the inria​​ mentoring program as mentor​​​‌ of one PhD student‌ from the centre Inria‌​‌ of Paris

11.2.2 Supervision​​

PY Oudeyer (co-)supervised the​​​‌ following PhD students:

• PhD‌ defended in 2025: Grgur‌​‌ Kovac, "Developmental training of​​ socio-cognitive abilities in AI​​​‌ systems", (supervisors:PF. Dominey and‌ PY. Oudeyer)
• PhD defended‌​‌ in 2025: Gauthier Hamon,​​ "Open-endedness in artificial life​​​‌ and articial intelligence: an‌ eco-evo-devo perspective" (supervisor: C.‌​‌ Moulin-Frier)
• PhD defended in​​ 2025: Nicolas Yax, "Studying​​​‌ cognitive and metacognitive skills‌ in foundation models" (supervisors:‌​‌ S. Palminteri, PY. Oudeyer)​​
• PhD defended in 2025:​​​‌ Clément Romac, "Grounding LLMs‌ with online RL", (supervisors:‌​‌ T. Wolf and PY.​​ Oudeyer) item PhD in​​​‌ progress: Julien Rosenberg, "Models‌ and experimental study of‌​‌ the co-development of curiosity​​ and metacognition in adolescents"​​​‌ (supervisors: H. Sauzéon, PY‌ Oudeyer)
• PhD in progress:‌​‌ Paul Tabbara, "Autotelic generative​​ AI systems for automated​​​‌ discovery in mathematics" (supervisors:‌ G. Baudart, PY. Oudeyer)‌​‌
• PhD in progress: Julien​​ Pourcel, "Autotelic LLMs that​​​‌ learn how to code",‌ (supervisors: C. Moulin-Frier and‌​‌ PY. Oudeyer)
• PhD in​​ progress: Thomas Carta, "LLM-based​​​‌ Autotelic deep reinforcement learning‌ agents", (supervisors: O. Sigaud,‌​‌ S. Lamprier and PY.​​ Oudeyer)
• PhD in progress:​​​‌ Jeremy Perez, "Studying mechanisms‌ and roles of curiosity‌​‌ in socio-cultural contexts" (supervisors:​​ C. Moulin-Frier, M. Derex,​​​‌ PY. Oudeyer)
• PhD in‌ progress: Timothé Boulet, "Controller‌​‌ synthesis for artificial agents​​ in simulated environments using​​​‌ generative AI" (supervisors C.‌ Moulin-Frier, X. Hinault, N.‌​‌ Fijalkow)
• PhD in progress:​​ Marko Cvjetko, "Autotelic exploration​​​‌ algorithms for automated search‌ of open-endedness in artificial‌​‌ life" (supervisors: C. Moulin-Frier,​​ PY. Oudeyer) item PhD​​​‌ in progress: Loris Gaven,‌ "Metacognitive prediction of learning‌​‌ progress for guiding autotelic​​ agents" (supervisors: PY. Oudeyer​​​‌ and C. Moulin-Frier)

H.‌ Sauzéon (co-)supervised the following‌​‌ PhD students:

• PhD defended​​ in 2025: M. POUPARD​​​‌ " Curious and thus‌ not overloaded !". (supervisors:H.‌​‌ sauzeon and A. tricot​​​‌ / CIFRE with CATIE)​
• PhD in progress: L.​‌ PETIOT " AR effect​​ on memory distorsions" (supervisors​​​‌ : H. sauzeon and​ P. Dragicevic) ( AEx​‌ IAM, 2023-25).
• PhD in​​ progress: C. DESVAUX "Design​​​‌ and Asssement of metacognitive​ interventions supporting curiosity and​‌ creativy at school" (Alloc.​​ MESRI, ED SP2).
• PhD​​​‌ in progress: J. DEYTS​ "Self-determination driven technologies for​‌ healthy aging" (Alloc. from​​ Projet ANR Innovcare)
• hD​​​‌ in progress: J. ROSENBERGER​ Curiosity-driven learnig as developmental​‌ function of metacognition in​​ adolescents aged of 12​​​‌ to 16 y/o. (supervisors​ : H. sauzeon and​‌ PY oudeyer(Alloc. from ORA​​ funds, ED SP2 -Univ.​​​‌ Bordeaux)
• PhD in progress:​ M. BOURDIL "A neurotechnological​‌ approach using EEG for​​ the characterising and the​​​‌ therapeutic treatment of smal​ vessels syndrome. (supervisors: F​‌ Lotte and H. sauzeon)​​ (Alloc from IHU-VBHI project)​​​‌

11.2.3 Juries

PY Oudeyer​ was a member of:​‌

• the selection committee of​​ the Inria Prizes from​​​‌ Académie des Sciences.
• the​ PhD juries of Marie​‌ Martin (Université Interdisciplinaire de​​ Paris), Théo Cachet (Sorbonne​​​‌ University) and J. Daly​ (Univ. Texas, Austin)
• the​‌ PhD "comité de suivi"​​ of Reem al Najjar​​​‌ (Sorbonne Université), Matthis Poupard​ (Univ. Bordeaux), Paul Pacaud​‌ (Université Paris Sciences Lettres)​​

Hélène Sauzéon was a​​​‌ part of 6 PhD​ boards :

• "Conception, développement​‌ et évaluation d'un exergame​​ en réalité augmentée pour​​​‌ la rééducation cognitivo-motrice d'enfants​ atteints de Paralysie Cérébrale​‌ ou de Lésions Cérébrales​​ Acquises : le projet​​​‌ TERAPACE by Maxime Balloufaud​ - Limoges
• "Optimizing sensory​‌ feedback and manual interaction​​ efficiency within XR experiments"​​​‌ by Julien Cauquis -​ Ecole nationale supérieure Mines-Télécom​‌ Atlantique Bretagne Pays de​​ la Loire
• "Careless or​​​‌ care-led innovation? : socio-ethnography​ of social robots and​‌ social tiesin eldercare settings​​ in France and Japan​​​‌ : tensions and contradictions​ in needs, temporalities and​‌ representations" by Yuko Tamaki​​ - Paris, EHESS
• "Neurocognitive​​​‌ mechanisms of self-referenced memory​ encoding: a naturalistic and​‌ embodied approach to episodic​​ memory" by Sylvain Penaud​​​‌ - Université Paris Cité​
• "Intrinsic vs. Extrinsic Motivation:​‌ Computational Modelling, Neural Bases,​​ and Clinical Applications" by​​​‌ Jade Seguin - Sorbonne​ université
• "Prise de décision​‌ lors de la planification​​ d'itinéraires avec des applications​​​‌ : une approche cognitive​ pour la régulation des​‌ flux voyageurs dans les​​ transports en commun" by​​​‌ Archana Prabhakar - Université​ Paris Cité

Cécile Mazon​‌ is permanent member of​​ the jury for Cognitive​​​‌ Sciences master thesis defenses​ (M1/M2) and for bachelor​‌ undergraduate projects (L3 MIASHS).​​

11.2.4 Support to public​​​‌ policies

PY. Oudeyer and​ H. Sauzéon and the​‌ whole team were involved​​ in several major actions​​​‌ to support public policies​ on the topic of​‌ AI and education. Members​​ of the team designed​​​‌ and conducted training sessions​ in different academies for​‌ supervisory staff and teachers,​​ e.g. ETAPP-IA day in​​​‌ Nouvelle-Aquitaine (January 2025); departmental​ training of CPE and​‌ documentary teachers of Nouvelle-Aquitaine​​ during a day at​​​‌ the Lycée Les Iris​ in Lormont (May 2025);​‌ Academic Days of Innovation​​ for teachers of Nouvelle-Aquitaine,​​​‌ Spring Days of Education​ Research at INSPEs, (June​‌ 2025); PhilosophIA Citizens' Convention​​ (April 2025), twin conference​​ of Cnesco/Cardie Charente-Maritime (January​​​‌ 2025), working group Education‌ and Cognitive Sciences of‌​‌ the academies of Créteil,​​ Versailles and Paris, scheduled​​​‌ for March 2026.

H.‌ Sauzéon and PY. Oudeyer‌​‌ were interviewed and wrote​​ reports to contribute to​​​‌ the report of French‌ Senate on AI and‌​‌ education.

PY Oudeyer​​ was auditioned by the​​​‌ commission on cultural and‌ educational affairs in the‌​‌ French parliament, to discuss​​ the major challenges and​​​‌ opportunities of AI and‌ education.

11.2.5 Educational and‌​‌ pedagogical outreach

Cécile Mazon​​ participated to events for​​​‌ promoting university programs in‌ cognitive science: the Salon‌​‌ de l’Étudiant (January 2025),​​ the University of Bordeaux​​​‌ Open Days (January 2025),‌ and the Orientation Days‌​‌ (May 2025).

11.3.1 Specific official responsibilities​​​‌ in science outreach structures‌

PY. Oudeyer collaborated with‌​‌ the Pix organization as​​ main scientific and editorial​​​‌ design consultant for the‌ Pix IA training modules,‌​‌ which will be dissemnated​​ to all French students​​​‌ in 4ème, 2nde and‌ CAP in 2026.

11.3.2‌​‌ Productions (articles, videos, podcasts,​​ serious games, ...)

PY.​​​‌ Oudeyer gave several public‌ talks on AI and‌​‌ education available on a​​ youtube channel.

D. Roy​​​‌ and P-Y. Oudeyer wrote‌ a popular science book‌​‌ to introduce generative AI​​ (mechanisms, applications, societal dimensions)​​​‌ to adolescents, as well‌ as to their teachers‌​‌ and families. It is​​ entitled "C'est (pas) moi,​​​‌ c'est l'IA", and was‌ published in september 2024‌​‌ by Nathan. It was​​ reviewed in widely distributed​​​‌ magazines (e.g. Magazine de‌ l'APEL) and radios (e.g.‌​‌ France Culture, RFI). The​​ web page of the​​​‌ book is here: link‌.

A. Torres-Leguet, C.‌​‌ Romac, T. Carta and​​ PY. Oudeyer produced the​​​‌ pedagogical video series "ChatGPT‌ explained in 5 mn",‌​‌ aimed at training generative​​ AI literacy in a​​​‌ wide diversity of students‌ (e.g. high school), available‌​‌ here: link. They​​ are under a Creative​​​‌ Commons licence, CC-BY, enabling‌ open and free reuse.‌​‌ They were already integrated​​ in the MOOC AI4T​​​‌ (link), as‌ well as in an‌​‌ internal training platform of​​ "Académie du Numérique du​​​‌ Ministère de la défense",‌ in a mobile app‌​‌ made by Inria with​​ educational materials related to​​​‌ AI (link),‌ and are being adapted‌​‌ and integrated in a​​ training platform for the​​​‌ whole population of civil‌ servants in France, coordinated‌​‌ by DINUM.

PY Oudeyer​​ wrote a note for​​​‌ the French educational institutions‌ on "IA générative, société‌​‌ et éducation: En quoi​​ l’IA générative représente-elle un​​​‌ enjeu dans la formation‌ des citoyens ?",‌​‌ in the context of​​ the Conférence de Consensus​​​‌ on Nouveaux Savoirs et‌ Nouvelles Compétences des Jeunes‌​‌ of Cnesco, (Nov. 2024)​​

Hélène Sauzéon wrote a​​​‌ web article on the‌ following topic: "Why agency‌​‌ is a key ability​​ in the workplace"

Hélène​​​‌ Sauzéon participated to the‌ "mental health and Technology"‌​‌ podcast organized by BPH​​ -Inserm (october 2025)​​​‌ in Bordeaux

Marie-Sarah Desvaux‌ walk interviewed by Curieux!‌​‌ Live on Educational Technologies​​ for learning

11.3.3 Participation​​​‌ in Live events

Jeremy‌ Perez and Clément Romac‌​‌ gave a presentation on​​​‌ Artificial Intelligence to high​ school teachers as part​‌ of the "Journée formation​​ IA pour les enseignant.es"​​​‌ at the Bordeaux INRIA​ Center on February 5th.​‌

Clément Romac gave a​​ talk on generative AIs​​​‌ to La main à​ la pâte, a French​‌ association promoting science in​​ classrooms.

Hélène Sauzéon ,​​​‌ Cécile Mazon , Sophie​ Lepennetier , Julien Rosenberger​‌ , Loris Gaven ,​​ Paul Tabbara and Julien​​​‌ Pourcel hosted a stand​ at the Village des​‌ Sciences on October 11th​​ and 12th. It gave​​​‌ an opportunity to introduce​ curiosity to visitors of​‌ CapScience, especially kids and​​ parents.

Hélène Sauzéon and​​​‌ Marie-Sarah Desvaux animated a​ workshop on curiosity-driven learning​‌ to teachers and trainers​​ during the "Learning Show"​​​‌ 2025 (13th of October)​ in Rennes

Hélène Sauzéon​‌ gave a talk at​​ the event organized by​​​‌ "Science with and for​ Society" by Université of​‌ Bordeaux: Samedis Sciences #4​​ "Artificial Intelligence and Education​​​‌ : the future of​ learning?"

Hélène Sauzéon gave​‌ a talk at the​​ event « Journée académique​​​‌ de l’expérimentation" organized by​ CARDIE Grenoble, Grenoble (14​‌ mai 2025)

Hélène Sauzéon​​ participated in "CoAnimation" for​​​‌ the Portes Fermées at​ INRIA, for a workshop​‌ to promote dialogue between​​ digital and social sciences​​​‌

Hélène Sauzéon participated in​ "Circuit scientifique Hors les​‌ murs" in October 2025​​

Hélène Sauzéon participated in​​​‌ the Chiche program, visiting​ 2 to 3 classrooms​‌

Marie-Sarah Desvaux gave a​​ talk on the use​​​‌ of Generative AI in​ classrooms to INSPE students​‌ of University of Bordeaux​​ (March 2025)

Marie-Sarah Desvaux​​​‌ animated a workshop on​ the use of Generative​‌ AI in classrooms during​​ the Journée Académique (August​​​‌ 2025) organized by CARDIE​ Poitiers

Marie-Sarah Desvaux gave​‌ an interactive talk on​​ curisoity-driven learning in classrooms​​​‌ to teachers during the​ Cogni'Forum 2025 (October) organized​‌ by "Apprendre et Former​​ avec les Sciences Cognitives"​​​‌

PY Oudeyer participated to​ several live events:

• (Feb​‌ 2025) Presentation of the​​ researcher job to high​​​‌ school students at La​ Sauque high-school (Nouvelle-Aquitaine)
• (March​‌ 2025) La créativité dans​​ tous ses états,​​​‌ with E. Koechlin, F.​ Guedy, P. Ribault, organized​‌ by Institute of Advanced​​ Studies, Paris.
• (Nov 2025)​​​‌ Presentation of the societal​ stakes of AI to​‌ a group of high-school​​ students in Compiègne, in​​​‌ the context of the​ Roberval prize event.
• (Jan​‌ 2025) D. Roy and​​ PY. Oudeyer, interview et​​​‌ présentation du livre "C'est​ (pas) moi, c'est l'IA"​‌, lors d'une rencontre​​ avec des collégiens organisée​​​‌ par la libraire Mollat.​
• (Feb. 2025) PY Oudeyer,​‌ Curiosité, cognition et intelligence​​ artificielle : Comment mieux​​​‌ apprendre à apprendre ?​, Série de conférence​‌ MIA Seconde.
• (Oct 2025)​​ PY Oudeyer, IA générative,​​​‌ société et éducation,​ for the conference on​‌ AI organized by PolarIA​​ at Fleurance, in the​​​‌ context of the annual​ science festival.

11.3.4 Others​‌ science outreach relevant activities​​

Press:

PY. Oudeyer was​​​‌ interviewed, or the work​ of the team was​‌ discussed, in various newspapers,​​ magazines and radios/podcasts:

• Espiloon​​​‌ (oct. 2025), IA, et​ maintenant les robots !​‌
• Telerama (Aout 2025), Le​​ risque, c’est l’affaiblissement de​​ la pensée: quand l’IA​​​‌ met l’école à l’épreuve‌
• Version Femina (Aout 2025),‌​‌ Comment leur apprendre à​​ utiliser l’IA?
• Télérama (Mai​​​‌ 2025), Intelligence artificielle :‌ comment bien accompagner les‌​‌ enfants ?
• Le Monde​​ (Feb. 2025), L’IA à​​​‌ l’école, une révolution déjà‌ en marche
• Magazine de‌​‌ l'APEL (Jan 2025) PY.​​ Oudeyer, L'IA, un outil​​​‌ pour la différenciation pédagogique‌, entretien réalisé pour‌​‌ le Magazing de l'APEL.​​

International journals​

• 34 articleG.Gautier​‌ Hamon, M.Mayalen​​ Etcheverry, B.-C. W.​​​‌Bert Wang-Chak Chan,​ C.Clément Moulin-Frier and​‌ P.-Y.Pierre-Yves Oudeyer.​​ Discovering Sensorimotor Agency in​​​‌ Cellular Automata using Diversity​ Search.Science Advances​‌ 11442025HAL​​DOIback to text​​​‌back to text
• 35​ articleE.Eric Meyer​‌, H.Hélène Sauzéon​​, I.Isabeau Saint-Supery​​​‌ and C.Cecile Mazon​. Evaluating a Web-Based​‌ Application to Facilitate Family-School-Health​​ Care Collaboration for Children​​​‌ With Neurodevelopmental Disorders in​ Inclusive Settings: Protocol for​‌ a Nonrandomized Trial.​​JMIR Research Protocols14​​​‌April 2025, e63378​HALDOIback to​‌ text
• 36 articleM.​​Marion Pech, M.​​​‌Maxime Adolphe, P.-Y.​Pierre-Yves Oudeyer and H.​‌Hélène Sauzéon. Broadening​​ the lens: A review​​​‌ of multi-object tracking task​ and its use in​‌ cognitive training.Acta​​ Psychologica258July 2025​​​‌, 105271HALDOI​back to text
• 37​‌ articleM.Matisse Poupard​​, F.Florian Larrue​​​‌, M.Martin Bertrand​, D.Dominique Liguoro​‌, H.Hélène Sauzéon​​ and A.André Tricot​​​‌. From Movement to​ Learning: Leveraging VR Behavioral​‌ Metrics to Evaluate Cognitive​​ Load and Curiosity.​​​‌International Journal of Human-Computer​ Studies209February 2026​‌, 103751HALDOI​​back to text
• 38​​​‌ articleM.Matisse Poupard​, F.Florian Larrue​‌, M.Martin Bertrand​​, D.Dominique Liguoro​​​‌, A.Andre Tricot​ and H.Hélène Sauzéon​‌. Using virtual reality​​ for enhancing neuroanatomy learning​​​‌ by optimizing cognitive load​ and intrinsic motivation..​‌Computers and Education235​​October 2025, 105332​​​‌HALDOIback to​ textback to text​‌
• 39 articleM.Matisse​​ Poupard, F.Florian​​​‌ Larrue, H.Hélène​ Sauzéon and A.André​‌ Tricot. A systematic​​ review of immersive technologies​​​‌ for education: effects of​ cognitive load and curiosity​‌ state on learning performance​​.British Journal of​​​‌ Educational Technology561​2025, 5-41HAL​‌DOIback to text​​back to text
• 40​​​‌ articleJ.Julien Pourcel​, C.Cédric Colas​‌ and P.-Y.Pierre-Yves Oudeyer​​. Self-Improving Language Models​​ for Evolutionary Program Synthesis:​​​‌ A Case Study on‌ ARC-AGI.Proceedings of‌​‌ Machine Learning Research2025​​HALDOIback to​​​‌ text
• 41 articleI.‌Isabeau Saint-Supery, H.‌​‌Hélène Sauzéon, E.​​Eric Meyer and C.​​​‌Cécile Mazon. CoEd,‌ an interactive website for‌​‌ the stakeholders of school​​ inclusion of children with​​​‌ ASD: an iterative design‌ including user testing.‌​‌Technology, Pedagogy and Education​​September 2025, 1-23​​​‌HALDOIback to‌ text
• 42 articleA.‌​‌Alexandr Ten, P.-Y.​​Pierre-Yves Oudeyer, M.​​​‌Michiko Sakaki and K.‌Kou Murayama. The‌​‌ Curious U : Integrating​​ Theories Linking Knowledge and​​​‌ Information-Seeking Behavior.Open‌ Mind9October 2025‌​‌, 1763-1785HALDOI​​back to text

International​​​‌ peer-reviewed conferences

• 43 inproceedings‌M. S.Mohamed Salim‌​‌ Aissi, C.Clément​​ Romac, T.Thomas​​​‌ Carta, S.Sylvain‌ Lamprier, P.-Y.Pierre-Yves‌​‌ Oudeyer, O.Olivier​​ Sigaud, L.Laure​​​‌ Soulier and N.Nicolas‌ Thome. Reinforcement Learning‌​‌ for Aligning Large Language​​ Models Agents with Interactive​​​‌ Environments: Quantifying and Mitigating‌ Prompt Overfitting.NAACL‌​‌ 2025 - Findings of​​ the Association for Computational​​​‌ LinguisticsAlbuquerque, United States‌Association for Computational Linguistics‌​‌May 2025, 7030-7046​​HALDOI
• 44 inproceedings​​​‌T.Timothé Boulet,‌ X.Xavier Hinaut and‌​‌ C.Clément Moulin-Frier.​​ Software Engineering Agents for​​​‌ Embodied Controller Generation :‌ A Study in Minigrid‌​‌ Environments.NeurIPS 2025​​ Efficient Reasoning WorkshopSan​​​‌ Diego (CA), United States‌December 2025HALback‌​‌ to textback to​​ text
• 45 inproceedingsM.​​​‌Marko Cvjetko, G.‌Gautier Hamon, C.‌​‌Clément Moulin-Frier and P.-Y.​​Pierre-Yves Oudeyer. Discovering​​​‌ and Controlling Diverse Self-Organised‌ Patterns in Cellular Automata‌​‌ Using Autotelic Reinforcement Learning​​.Alife 2025 -​​​‌ Conference on Artificial Life‌Kyoto, JapanOctober 2025‌​‌HALback to text​​back to text
• 46​​​‌ inproceedingsJ.Juliette Deyts‌, R.Rafik Belloum‌​‌, L.Lucile Dupuy​​, L.Loïc Blouin​​​‌ and H.Hélène Sauzéon‌. Towards an Interactive‌​‌ Longitudinal Visualization of Daily​​ Living Activities for Aging​​​‌ in Place.IHM‌ 2025 - 36e Conférence‌​‌ Internationale Francophone sur l’Interaction​​ Humain-MachineToulouse, FranceNovember​​​‌ 2025HALback to‌ text
• 47 inproceedingsL.‌​‌Loris Gaven, T.​​Thomas Carta, C.​​​‌Clément Romac, C.‌Cédric Colas, S.‌​‌Sylvain Lamprier, O.​​Olivier Sigaud and P.-Y.​​​‌Pierre-Yves Oudeyer. MAGELLAN:‌ Metacognitive predictions of learning‌​‌ progress guide autotelic LLM​​ agents in large goal​​​‌ spaces.ICML 2025‌ - 42nd International Conference‌​‌ on Machine Learning267​​Vancouver (BC), Canada2025​​​‌HALback to text‌back to textback‌​‌ to text
• 48 inproceedings​​S.Sina Khajehabdollahi,​​​‌ G.Gautier Hamon,‌ M.Marko Cvjetko,‌​‌ P.-Y.Pierre-Yves Oudeyer,​​ C.Clément Moulin-Frier and​​​‌ C.Cédric Colas.‌ Expedition &amp; Expansion: Leveraging‌​‌ Semantic Representations for Goal-Directed​​ Exploration in Continuous Cellular​​​‌ Automata.ALIFE 2025‌ - Conference on Artificial‌​‌ LifeKyoto, Japan2025​​HALback to text​​​‌back to text
• 49‌ inproceedingsG.Grgur Kovač‌​‌, J.Jérémy Perez​​​‌, R.Rémy Portelas​, P. F.Peter​‌ Ford Dominey and P.-Y.​​Pierre-Yves Oudeyer. Recursive​​​‌ Training Loops in LLMs:​ How training data properties​‌ modulate distribution shift in​​ generated data?Proceedings of​​​‌ the 2025 Conference on​ Empirical Methods in Natural​‌ Language ProcessingEMNLP 2025​​ - Conference on Empirical​​​‌ Methods in Natural Language​ ProcessingSuzhou, ChinaAssociation​‌ for Computational LinguisticsNovember​​ 2025, 32278-32297HAL​​​‌DOIback to text​back to text
• 50​‌ inproceedingsA.Aleksa Marusic​​, S. M.Sao​​​‌ Mai Nguyen and A.​Adriana Tapus. Skeleton-Based​‌ Transformer for Classification of​​ Errors and Better Feedback​​​‌ in Low Back Pain​ Physical Rehabilitation Exercises.​‌ICORR 2025 - 19th​​ IEEE/RAS-EMBS International Conference on​​​‌ Rehabilitation RoboticsHRI '23:​ Companion of the 2023​‌ ACM/IEEE International Conference on​​ Human-Robot InteractionMichigan, United​​​‌ StatesIEEEMay 2025​HAL
• 51 inproceedingsT.​‌Thomas Michel, M.​​Marko Cvjetko, G.​​​‌Gautier Hamon, P.-Y.​Pierre-Yves Oudeyer and C.​‌Clément Moulin-Frier. Exploring​​ Flow-Lenia Universes with a​​​‌ Curiosity-driven AI Scientist: Discovering​ Diverse Ecosystem Dynamics.​‌Artificial Life Conference Proceedings​​ 37ALIFE 2025 -​​​‌ Conference on Artificial Life​20251Kyoto /​‌ Virtual, Japan2025,​​ 68HALDOIback​​​‌ to textback to​ text
• 52 inproceedingsB.​‌Bastien Morel, C.​​Clément Moulin-Frier and P.​​​‌Pascal Barla. Complex​ System Exploration with Interactive​‌ Human Guidance.WIVACE​​ 2025 - XIX International​​​‌ Workshop on Artificial Life​ and Evolutionary ComputationSiena,​‌ Italy2025HALback​​ to text
• 53 inproceedings​​​‌L.Léana Petiot,​ H.Hélène Sauzéon and​‌ P.Pierre Dragicevic.​​ The Effect of Augmented​​​‌ Reality on Involuntary Autobiographical​ Memory.CHI 2025​‌ - Conference on Human​​ Factors in Computing Systems​​​‌Yokohama, JapanApril 2025​HALDOIback to​‌ textback to text​​
• 54 inproceedingsL.Léana​​​‌ Petiot, H.Hélène​ Sauzéon and P.Pierre​‌ Dragicevic. Using Visual​​ Cues to Prevent Memory​​​‌ Confusion Between the Virtual​ and the Real in​‌ Augmented Reality.Late-breaking​​ work / poster published​​​‌ in the CHI extended​ abstracts (CHI EA ’25)​‌CHI 2025 - Conference​​ on Human Factors in​​​‌ Computing SystemsYokohama, Japan​April 2025HALDOI​‌back to text
• 55​​ inproceedingsBest paperM.​​​‌Max Taylor-Davies, G.​Gautier Hamon, T.​‌Timothé Boulet and C.​​Clément Moulin-Frier. Emergent​​​‌ kin selection of altruistic​ feeding via non-episodic neuroevolution​‌.Applications of Evolutionary​​ ComputationInternational Conference on​​​‌ the Applications of Evolutionary​ Computation (Part of EvoStar)​‌Trieste, Italy2025HAL​​back to textback​​​‌ to textback to​ text
• 56 inproceedingsN.​‌Nicolas Yax, P.-Y.​​Pierre-Yves Oudeyer and S.​​​‌Stefano Palminteri. PhyloLM​ : Inferring the Phylogeny​‌ of Large Language Models​​ and Predicting their Performances​​​‌ in Benchmarks.ICLR​ 2025Singapore, Singapore2025​‌HALback to text​​

National peer-reviewed Conferences

• 57​​​‌ inproceedingsS.Sofiya Kobylyanskaya​, C.Catherine de​‌ Vulpillères and P.-Y.Pierre-Yves​​ Oudeyer. A hybrid​​​‌ AI approach to educational​ technologies : augmenting ITS​‌ with generative AI.​​Actes de l'atelier Intelligence​​ Artificielle générative et ÉDUcation​​​‌ : Enjeux, Défis et‌ Perspectives de Recherche 2025‌​‌ (IA-ÉDU)20e Conférence en​​ Recherche d’Information et Applications​​​‌ (CORIA) 32ème Conférence sur‌ le Traitement Automatique des‌​‌ Langues Naturelles (TALN) 27ème​​ Rencontre des Étudiants Chercheurs​​​‌ en Informatique pour le‌ Traitement Automatique des Langues‌​‌ (RECITAL) Les 18e Rencontres​​ Jeunes Chercheurs en RI​​​‌ (RJCRI)Marseille, FranceATALA‌ et ARIA2025,‌​‌ 145-148HAL

Conferences without​​ proceedings

• 58 inproceedingsH.​​​‌Hana Al-Mrayati, É.‌Éric Meyer, H.‌​‌Hélène Sauzéon and C.​​Cécile Mazon. Value​​​‌ of Parent-Professional partnerships in‌ inclusive education : evaluation‌​‌ of a digital tool​​ supporting Family-School-Healthcare collaboration for​​​‌ students with NDD.‌EIAH 2025 - 12ème‌​‌ Conférence sur les Environnements​​ Informatiques pour l'Apprentissage Humain​​​‌Villeneuve d’Ascq, FranceJune‌ 2025HALback to‌​‌ text
• 59 inproceedingsG.​​Guillaume Levy, C.​​​‌Cédric Colas, P.-Y.‌Pierre-Yves Oudeyer, T.‌​‌Thomas Carta and C.​​Clément Romac. WORLDLLM:​​​‌ IMPROVING LLMS' WORLD MODELING‌ USING CURIOSITY-DRIVEN THEORY-MAKING.‌​‌RLDM 2025 - Multi-disciplinary​​ Conference on Reinforcement Learning​​​‌ and Decision MakingDublin,‌ IrelandJune 2025HAL‌​‌back to textback​​ to text
• 60 inproceedings​​​‌J.Jérémy Perez,‌ G.Grgur Kovač,‌​‌ C.Corentin Léger,​​ C.Cédric Colas,​​​‌ G.Gaia Molinaro,‌ M.Maxime Derex,‌​‌ P.-Y.Pierre-Yves Oudeyer and​​ C.Clément Moulin-Frier.​​​‌ When LLMs Play the‌ Telephone Game: Cultural Attractors‌​‌ as Conceptual Tools to​​ Evaluate LLMs in Multi-turn​​​‌ Settings.The Thirteenth‌ International Conference on Learning‌​‌ Representations (ICLR 2025)Singapour,​​ Singapore2025HALback​​​‌ to text

Doctoral dissertations‌ and habilitation theses

• 61‌​‌ thesisG.Gautier Hamon​​. Towards open-ended dynamics​​​‌ in Artificial Life and‌ Artificial Intelligence : an‌​‌ eco-evo-devo perspective.Université​​ de BordeauxMarch 2025​​​‌HAL
• 62 thesisG.‌Grgur Kovač. Building,‌​‌ evaluating and understanding socio-cultural​​ AI : leveraging concepts​​​‌ and methods from human‌ sciences.Université de‌​‌ BordeauxNovember 2025HAL​​
• 63 thesisM.Matisse​​​‌ Poupard. Curious and‌ therefore not overloaded: Towards‌​‌ an integrated understanding of​​ curiosity and cognitive load​​​‌ in XR learning environments.‌.Université de bordeaux‌​‌September 2025HAL

Reports​​ & preprints

• 64 misc​​​‌J.Jérémy Perez,‌ M.Maxime Derex,‌​‌ P.-Y.Pierre-Yves Oudeyer and​​ C.Clément Moulin-Frier.​​​‌ Intrinsic motivation is key‌ to understanding peer cultures:‌​‌ Commentary on target article:​​ Lew-Levy, S. &amp; Amir,​​​‌ D. (2024). Children as‌ agents of cultural adaptation.‌​‌ Behavioral and Brain Sciences,​​ 1–68.November 2025​​​‌HALback to text‌
• 65 miscM.Matisse‌​‌ Poupard, F.F​​ Larrue, M.Martin​​​‌ Bertrand, D.D‌ Liguoro, H.H‌​‌ Sauzéon and A.A​​ Tricot. Enhancing Anatomy​​​‌ Learning Through Mixed Reality-Supported‌ Drawing: Investigating Learning Performance,‌​‌ Cognitive Load, and Intrinsic​​ Motivation.April 2025​​​‌HALback to text‌
• 66 miscM.Matisse‌​‌ Poupard, F.Florian​​ Larrue, M.Martin​​​‌ Bertrand, D.Dominique‌ Liguoro, H.Hélène‌​‌ Sauzéon and A.André​​ Tricot. Reducing Load,​​​‌ Fostering Curiosity: Empirical Validation‌ of the IMCLM-XR.‌​‌July 2025HALback​​​‌ to text

Scientific popularization​

• 67 miscC.Clément​‌ Romac, P.-Y.Pierre-Yves​​ Oudeyer and T.Thomas​​​‌ Carta. Can AIs​ understand our world? Functionally​‌ grounding LLMs in interactive​​ environments..April 2025​​​‌HAL