How can artificial agents discover skills, invent ways to communicate, and develop complex cooperative strategies without being told what to do? ECOCURL studies how curiosity, learning and evolution in populations of interacting agents can drive the spontaneous emergence of shared communication and collective abilities, echoing the dynamics that may have shaped language and culture in living populations.
Today's powerful artificial intelligence systems — from game-playing agents to language models — are largely designed: their skills are shaped by human-defined objectives, large curated datasets and engineered reward signals. Yet many of the most remarkable abilities found in nature — language, technological skills, complex social organization — were not designed: they emerged in populations of living beings interacting with their environment and with each other. ECOCURL set out to investigate to what extent similar dynamics can be reproduced in artificial agents, with the long-term ambition of designing AI systems that autonomously discover and master new skills through their interactions, rather than being told what to do.
The project addressed three core questions. First, can populations of artificial agents driven by curiosity — the internal drive to explore and master new abilities — collectively discover a rich diversity of skills through their interactions? Second, what kinds of cooperation and communication systems emerge between such agents, and how does the structure of their environment and of their internal cognitive architecture shape these systems? Third, how can the resulting shared communication systems support the discovery of increasingly complex cooperative strategies, opening the way to forms of cultural evolution within artificial populations?
Answering these questions required bringing together several lines of research that had remained largely separate. Curiosity-driven learning, originally developed for individual agents discovering varied skills autonomously, had not been transposed to populations. The study of how shared communication systems can emerge between agents had been confined to relatively simple settings. The dynamics of mixed cooperation and competition, known to produce open-ended sequences of increasingly complex behaviors, had remained mostly disconnected from communication research. ECOCURL combined these strands to study how curiosity, communication and competition can jointly fuel the open-ended discovery of skills and collective behaviors at the population level.
The questions raised by the project naturally extend beyond multi-agent learning. They connect to ecological and evolutionary dynamics, since the conditions of an environment — its complexity, its variability, its distribution of resources — strongly shape which behaviors emerge in a population. They connect to cultural evolution, since populations capable of inventing and transmitting skills and signals are themselves cultural systems. They also connect to automated scientific discovery: the curiosity-driven algorithms developed for artificial agents turn out to be powerful tools for exploring complex dynamical systems, suggesting that AI inspired by living curiosity can also serve as an instrument for studying nature itself.
The methodology of ECOCURL consisted in building artificial worlds in which populations of agents can learn, interact and evolve, then observing which collective behaviors and communication patterns emerge from their interactions. The complexity of these worlds was progressively increased throughout the project, from small grid-based environments with a few agents up to massively populated simulations involving thousands of agents.
At the heart of the project, agents were equipped with reinforcement learning — a family of algorithms in which an agent improves its behavior through trial and error guided by rewards. Building on this foundation, agents were given the ability to set their own goals and to be driven by curiosity to discover and master new skills, drawing on previous work in autonomous goal-directed exploration. The compositional structure of these goals — the way simple skills can be combined into more elaborate ones — provided a basis for studying how shared symbolic communication can emerge between agents engaged in cooperative tasks.
To address questions that go beyond individual learning, the methodological toolkit was progressively extended. Artificial evolution was used to study how learning architectures and behaviors are shaped, over generations, by ecological pressures such as resource scarcity, environmental variability or kinship. Techniques originally developed to train agents on wide distributions of situations were adapted to multi-agent settings, allowing artificial populations to acquire cooperative strategies that generalize to previously unseen tasks. Large language models were introduced as a new experimental tool: by playing the role of communicating agents within controlled simulations, they enabled the study of cultural evolution and information transmission in populations endowed with rich, compositional language.
These ingredients converged in the development of Vivarium, a flexible simulation platform designed to host populations of thousands of interacting agents in a continuous environment with realistic physics. Built to exploit modern graphics processors, Vivarium can be used at multiple levels — from interactive web demonstrations accessible without programming, to large-scale supercomputing experiments. Finally, the same family of curiosity-driven exploration algorithms developed for artificial agents was applied to a different kind of question: the systematic exploration of complex computational worlds — such as cellular automata, in which intricate patterns emerge from simple local rules — opening a new direction in which AI methods themselves serve as instruments for scientific discovery.
A central finding of the project concerns the connection between curiosity and communication. When artificial agents are each driven by their own curiosity, they independently select different goals — a powerful driver of skill diversity, but also an obstacle to cooperation, since two agents pursuing different goals cannot coordinate. The project showed that this tension is naturally resolved when agents can communicate: a shared signaling protocol spontaneously emerges as a way to align goals, allowing decentralized populations to match the cooperation levels achieved under centralized training (Nisioti et al., 2023). Communication thus appears not as an added feature but as a solution to the coordination problem created by curiosity itself. In environments combining many possible activities, such curiosity-driven populations also acquired cooperative strategies that generalized far beyond their training distribution.
A second line of results concerns the role of the environment in shaping behaviors. A general computational framework was developed for long-running evolutionary simulations in which artificial agents reproduce, die and adapt without resetting their world (Hamon et al., 2023). In such simulations, altruistic resource transfer between generations spontaneously appeared under conditions of kin recognition, and ecological niches were found to favor adaptability and behavioral plasticity even in stable environments. Populations exposed to fires propagating across a landscape further evolved collective strategies that suppress environmental extremes while sustaining their resource base (Sánchez-Fibla et al., 2024).
On the cultural side, the project showed that the way agents share information across a population strongly affects what the population can invent Nisioti et al., 2022. Populations alternating between individual exploration and group-level sharing outperformed fully-connected populations on tasks requiring the gradual combination of innovations, mirroring patterns observed in human laboratory experiments. Experiments with populations of large language models passing messages along chains revealed cultural attractors in the evolution of text — biases towards specific levels of toxicity, positivity, length and difficulty (Perez et al., 2025) — and an open-source framework was released to study these dynamics in artificial populations.
Finally, the curiosity-driven exploration methods developed for artificial agents proved to be powerful instruments for scientific discovery in their own right. Applied to cellular automata — simulated worlds where complex patterns emerge from simple local rules — these methods automated the search for self-organized structures with life-like properties, including primitive forms of agency and ecosystems of interacting species (Hamon et al., 2025). The methodology was packaged in open-source software now used to study other complex systems, including biological networks, in collaboration with international partners.
Beyond the scientific results described above, the project produced concrete tools and community outcomes. Foremost is Vivarium, an open-source simulation platform hosting populations of thousands of artificial agents that learn, communicate and evolve in a continuous physics-based environment. It serves a wide range of audiences: students can interact with it through a browser-based interface without writing code, while researchers run massively parallel simulations on graphics processors. Vivarium has been adopted as a teaching platform at Pompeu Fabra University and in Bordeaux engineering schools. The project also organized academic workshops on emergent communication (SMILES) and community hackathons (Hack1Robo) gathering students, researchers and companies, with several follow-up start-ups.
The project also opened two new research axes that anchor ongoing work. The eco-evolutionary line — studying how environmental and ecological dynamics shape the behavior of agent populations — produced findings on the spontaneous emergence of altruistic behaviors under kin recognition (Taylor-Davies et al., 2025, best paper and best student paper awards at EvoStar 2025), and led to an industrial collaboration with Pontos on the sustainable management of marine ecosystems. A second axis on cultural evolution emerged through experiments using large language models as populations of communicating agents — a new experimental paradigm to study how information, norms and innovations propagate in artificial populations, with direct relevance for human–AI ecosystems.
A complementary research line developed during the project applied its curiosity-driven algorithms to the systematic exploration of complex computational systems. This line attracted broader public attention, including a feature article in Pour la Science on how artificial intelligence can study the emergence of life-like dynamics (Moulin-Frier et al., 2024), and international awards including the best paper award at the ALife conference (Plantec et al., 2023) and a prize at the Virtual Creature Competition.
Several long-term research directions emerge from these results. One concerns the design of communities mixing humans and curiosity-driven AI agents — both for understanding the cultural dynamics of hybrid ecosystems in which humans and large language models exchange information, and for organizing AI agents that collaborate with human researchers on scientific exploration. A long-term challenge is to integrate the project's contributions into a unified framework in which biological evolution, social interactions and cultural transmission jointly produce open-ended forms of intelligence — building on the conceptual framework developed in the principal investigator's habilitation thesis (Moulin-Frier, 2022).