Generative AI Collective Behaviour: New Interactionist Paradigm Advances Understanding of Multi-Generative Systems

Understanding the collective behaviour of large language models (LLMs) represents a critical challenge, with far-reaching consequences for society. Laura Ferrarotti from Fondazione Bruno Kessler, Gian Maria Campedelli of the University of Trento, and Roberto Dessì from Not Diamond, alongside Andrea Baronchelli et al., propose a new interactionist paradigm to systematically examine how pre-existing knowledge and inherent values within LLMs combine with social context. This research highlights the unique characteristics of these models , their extensive pre-training and capacity for in-context learning , demanding a shift in theoretical foundations and analytical tools. By outlining four key directions for development, the authors aim to improve both understanding and responsible deployment of LLM-based collective systems.

Scientists Background

Scientists demonstrate a critical need to understand the collective behaviour of large language model (LLM)-based agents, revealing implications for both the risks and benefits these systems pose to society. This research establishes that the unique characteristics of LLMs, extensive pre-trained knowledge, inherent social biases, and adaptive in-context learning, necessitate a new interactionist paradigm for systematically examining how prior knowledge and embedded values shape emergent phenomena in multi-agent generative AI systems. The team achieved a conceptual framework built upon four key pillars to facilitate the development and deployment of LLM-based collectives, focusing on theoretical foundations, methodological tools, and trans-disciplinary collaboration. The study unveils a significant gap in current multi-agent system (MAS) research, which traditionally relies on reinforcement learning frameworks ill-suited to LLMs’ pre-existing knowledge base.

Unlike agents learning from scratch, LLM-based agents arrive with substantial pre-trained knowledge and implicit social priors, demanding alternative theoretical approaches to understand their collective behaviour. Experiments show that understanding how this pre-existing knowledge interacts with the social context of multi-agent interactions is crucial for predicting and managing emergent behaviours, norms, and biases within these systems. This work proposes an interactionist paradigm, drawing parallels with longstanding debates in human behaviour, to examine the interplay between individual agent traits and dynamic interaction. The research establishes four core components: an interactionist theory to illuminate determinants of collective behaviour, causal inference for developing safe and functional systems, information theory as a quantitative language for studying knowledge distribution, and a new field of ‘sociology of machines’ for testing theoretical and empirical frameworks.

By integrating these disciplines, scientists aim to move beyond analysing individual agents or isolated environmental factors, instead focusing on the complex interplay that drives emergent collective behaviours. The proposed paradigm offers a comprehensive framework for grounding the design and training of Gen-AI agents, addressing phenomena that cannot be fully explained by examining agents in isolation. Its relevance extends beyond the immediate concerns of alignment and safety within MAS, contributing to a broader understanding of complex emergent behaviours. Notably, the team argues that this interactionist approach will remain valuable even with future advancements in AI technology, as pre-trained knowledge is likely to remain a defining characteristic of intelligent agents.

LLM Collectives Learn Through Interaction and Adaptation

This work details a new interactionist paradigm for understanding the collective behaviour of large language models (LLMs), moving beyond individual agent performance to examine emergent phenomena in multi-generative systems. Scientists propose a framework centered on learning through interaction, drawing parallels with social and cultural learning observed in natural systems. The research highlights that adaptive strategies frequently emerge through self-organization and real-time learning, processes significantly enhanced by interaction with others. Experiments reveal a conceptual shift from simple Gen-AI agents, initially trained via pre-training, supervised fine-tuning, and reinforcement learning from human feedback, towards collectives of agents engaged in interactive in-context learning (ICL) tasks.

The team modelled this progression, demonstrating how individual behaviours, established through initial training, transition into complex collective behaviours when agents interact. This interaction is formalized through a learning function, fi, mapping social information (Di) and environmental observations (Ei) to internal states (Hi), and an information exchange function, φt, governing data flow between agents. Measurements confirm that each model, Mi, updates its hypothesis hi at each time step t according to the equation h(t+1) i = fi(e(t+1) i, d(t+1) i ), where d(t+1) represents the data received from other agents and e(t+1) i denotes the agent’s environmental observation. The study emphasizes that this dynamic exchange, defined by φt(h(t) 1, ., h(t) n ), creates a constantly shifting environment, driving innovation and adaptation as agents respond to each other’s evolving strategies.

Researchers posit that interactive learning embodies an evolutionary feedback loop, where agents refine behaviours through socially mediated signals, leading to both individual adaptation and the transmission of collective norms. This framework considers a system of n artificial models, each contributing to and receiving information, ultimately shaping the collective’s ability to navigate complex environments. The work suggests that understanding these interactions is crucial for developing and deploying LLM-based collectives effectively.