Source-linked AI summary
Generative AI collective behavior needs an interactionist paradigm
Laura Ferrarotti, Gian Maria Campedelli, Roberto Dessì, Andrea Baronchelli, Giovanni Iacca, Kathleen M. Carley, Alex Pentland, Joel Z. Leibo, James Evans, Bruno Lepri
TL;DR
The paper addresses how to understand collective behavior in LLM-based agents whose pre-trained knowledge, social priors, and in-context adaptation shape interactions. It proposes an interactionist paradigm combining transdisciplinary theory and causal methods, and concludes that interactive collectives can generate context-dependent social behaviors requiring dedicated study. The paper also notes that emergent behaviors may be difficult to interpret or persist beyond their original context.
Problem
LLM collectives differ from tabula rasa MARL agents because pre-trained knowledge and social priors affect interaction, while existing benchmarks provide little insight into these effects.
Method
The paper proposes an interactionist paradigm integrating sociological theory, causal identification, information-theoretic analysis, and transdisciplinary perspectives for studying collective machine behavior.
Results
Interactive ICL can foster social behaviors such as linguistic norms, collective biases, and shifting conventions under environmental and experimental conditions.
Takeaways & Limitations
Understanding and guiding LLM collectives requires frameworks that examine both individual agents and the social systems arising from their interactions.
Takeaways & Limitations
Emergent behaviors may be difficult to interpret or control, and context-dependent behaviors may disappear or evolve when contextual information is not transferred.
Abstract
from arXiv · showhide
In this article, we argue that understanding the collective behavior of agents based on large language models (LLMs) is an essential area of inquiry, with important implications in terms of risks and benefits, impacting us as a society at many levels. We claim that the distinctive nature of LLMs--namely, their initialization with extensive pre-trained knowledge and implicit social priors, together with their capability of adaptation through in-context learning--motivates the need for an interactionist paradigm consisting of alternative theoretical foundations, methodologies, and analytical tools, in order to systematically examine how prior knowledge and embedded values interact with social context to shape emergent phenomena in multi-agent generative AI systems. We propose and discuss four directions that we consider crucial for the development and deployment of LLM-based collectives, focusing on theory, methods, and trans-disciplinary dialogue.
1 Introduction
LLM-based agents are increasingly interacting in collectives, creating emergent behaviors with societal risks and benefits that individual-agent analysis cannot fully explain. The paper proposes an interactionist paradigm integrating theory, causal methods, information theory, and sociology to study how pre-trained knowledge and social context shape these phenomena.
- LLM-based agents are moving from isolated operation toward increasingly autonomous interaction across diverse domains.
- Collective behavior requires systematic investigation because interactions can produce societal benefits and risks, motivating new benchmarks, evaluation methods, and theoretical tools.
- Unlike tabula rasa multi-agent reinforcement-learning systems, LLM agents bring substantial pre-trained knowledge and rich social priors into interaction.
- The proposed interactionist paradigm examines how pre-trained knowledge, embedded values, individual traits, and social behavior interact to produce collective phenomena.
- Its four pillars are interactionist theory, causal inference, information theory, and a sociology of machines that moves beyond human-centric social assumptions.
- The framework combines AI agency, causal inference, information theory, and sociological perspectives because emergent machine behavior cannot be explained by agents or environments in isolation.
2 Learning through interaction
Interactive learning treats agents as mutually engaged participants that adapt through socially mediated information, rather than learning in isolation. In AI collectives, this can support distributed adaptation and cooperation while also enabling harmful behaviors, cascading failures, and accountability challenges.
- Interactive learning requires agents to be concurrently engaged in a shared social situation as both sources and receivers of information.
- Social interaction can produce adaptive feedback loops in which agents refine behavior through signals, rewards, or punishments and stabilize collective norms.
- An interactive learning system comprises independent models that map social information and environmental observations to internal hypotheses, which they update through interaction.
- Distributed information enables collective problem-solving when individual agents lack complete information, while cooperation supports task specialization and performance in collaborative environments.
- Interactive agents may develop competencies, transfer norms, and adapt to changing environments, but interconnected systems can also propagate deviant behavior, errors, misinformation, and harmful strategies.
- Networked collectives can become difficult to interpret and control, complicating causal inference and accountability as failures or harmful actions spread across agents.
3 Learning socially: from MARL to collectives of Gen-AI agents
MARL studies social learning as an emergent capability of generic learning dynamics, whereas Gen-AI collectives bring pre-trained knowledge, social priors, and in-context adaptation into interaction. This difference motivates new constructs and benchmarks for evaluating emergent coordination, communication, and knowledge use.
- MARL and social learning: Generic reinforcement learning can produce social learning behaviors without specialized imitation mechanisms or built-in social inductive biases.Agents may acquire imitation and observational learning when environments provide useful information through other agents.
- MARL and social learning: MARL research reports social learning phenomena including copying, neighbor-based learning, cultural transmission, and cultural accumulation.These behaviors can emerge through balancing individual and social learning in interactive environments.
- Learning phases: LLM agents combine pre-training, supervised fine-tuning, RLHF, and interactive in-context learning, with ICL enabling adaptation from prompt context without weight updates.The pipeline establishes general knowledge, task-specific skills, aligned behavior, and adaptive deployment behavior.
- Collectives of Gen-AI agents: Multi-agent interactive ICL can foster linguistic norms, collective biases, convention shifts, self-organization, persuasion, negotiation, cooperation, competition, and coordination.The resulting behaviors depend on conditions such as assigned roles, contexts, and objectives.
- Collectives of Gen-AI agents: In-context adaptation can be difficult to steer because crafting prompts that elicit desirable agent behavior remains challenging.Natural-language prompts may make specification more accessible, but desirable behavior is not guaranteed.
- Evaluation: Because Gen-AI collectives reuse pre-interaction knowledge and conversational context, standard MARL reward metrics cannot capture their emergent coordination, communication quality, and world-knowledge use.The paper calls for new theoretical constructs and benchmarking protocols tailored to these properties.
4 An interactionist paradigm to study generative agents’ collective behaviors
The paper proposes an interactionist paradigm that explains collective behavior through the interplay of agents’ internalized priors and social situations. It combines interactionist theory, causal inference, information theory, and a sociology of machines to study and guide emergent machine behavior.
- Interactionist theory: The proposed paradigm examines collective behavior through both individual traits and emergent interaction dynamics.It addresses how pre-trained priors and social contexts jointly shape group outcomes.
- Interactionist theory: Interactionist theory treats Gen-AI collective behaviors as adaptive mechanisms shaped by internalized priors and interactive, in-context social learning.The framework also asks how interactions may influence agents’ traits and priors in later contexts.
- Causal inference: Causal inference is proposed to identify causal pathways governing emergent phenomena that cannot be explained by single-agent behavior alone.The paper calls for benchmarks, evaluation protocols, and experimental designs tailored to causal questions in multi-agent systems.
- Sociology of machines: A sociology of machines would study how roles, incentives, interaction structures, and learning dynamics produce coordination, conflict, norm formation, and deviance.The authors argue that human sociology provides useful foundations but cannot fully explain autonomous AI collectives.
5 Alternative Views
The paper positions its interactionist paradigm as complementary to critiques of LLM-based multi-agent systems and accounts of behavioral uniformity. It focuses on how agents with prior representations behave in social contexts and how collective phenomena emerge across implementations.
- Alternative views: The paradigm addresses how agents with prior representations behave in social contexts, rather than prescribing a specific architecture or training regime.It examines influence, information propagation, and collective emergence across implementations.
- Alternative views: Observed social conventions, polarization, and coordination patterns challenge the claim that shared training necessarily produces uniform collective behavior.The paper notes that even identically pretrained agents can produce diverse collective outcomes.
- Alternative views: Classical MARL and Gen-AI collectives differ because generative agents may vary in fine-tuning, prompting, retrieval context, interaction history, and architecture.These differences make both initial conditions and contextual factors relevant to emergent behavior.
- Alternative views: Current agent failures remain useful for theory development because explaining them requires causal attribution, information-flow analysis, and identification of coordination failures.The proposed paradigm targets mechanisms driving collectives independently of individual-agent sophistication.
6 Conclusions
The paper argues that LLM collectives require a new, transdisciplinary framework because pre-trained knowledge, social priors, and in-context adaptation generate behaviors existing MARL tools cannot fully explain or manage. It identifies four research directions for systematic evaluation and steering.
- Conclusions: LLM collectives pose an urgent societal research challenge because their pre-trained knowledge, implicit social priors, and in-context adaptation generate complex emergent behaviors.The paper argues that current MARL-oriented theoretical tools are not sufficient to fully explain or manage these behaviors.
- Conclusions: The proposed framework draws on cognitive science, social and cultural learning theory, machine learning, and transdisciplinary dialogue.Its purpose is to support systematic study of collective behavior across theory, methods, and deployment.
- Research directions: The paper proposes interactionist benchmarks that isolate situation factors such as prompts and interaction history from person factors such as priors, scale, and alignment.These benchmarks are intended to separate contextual and model-internal contributions to behavior.
- Research directions: The proposed methods include causal identification strategies, information-theoretic measures of influence and consensus, and empirical sociology using collectives as controlled model organisms.Together these directions target causal pathways, information flow, and social processes in machine collectives.
- Conclusions: Such a framework could help anticipate and steer emergent dynamics toward beneficial cooperation, reduced risk, and alignment with human values.The stated uses include both fostering beneficial outcomes and mitigating harmful dynamics.