Learning to Make Friends: Coaching LLM Agents toward Emergent Social Ties

Philipp J. Schneider, Lin Tian, Marian-Andrei Rizoiu · 2025

Read Learning to Make Friends: Coaching LLM Agents toward Emergent Social Ties by Philipp J. Schneider, Lin Tian, Marian-Andrei Rizoiu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-19299.txt.

Introduction

Philipp J Schneider, Lin Tian, and Marian-Andrei Rizoiu ask whether language model agents can learn to form social ties the way people do online. Human networks grow through likeness, mutual exchange, and social approval, and it is unclear which memory and learning mechanisms let artificial agents reproduce those patterns.

Their simulation lets agents interact repeatedly, judge each other, and adapt through in-context learning sped by a coaching signal. Reward functions mirror real motives such as seeking information, self-presentation, coordination, and emotional support. Watching network structures emerge from those rewards helps researchers study online behavior without experimenting on real users.

Catalog record

CREATORPhilipp J. Schneider, Lin Tian, Marian-Andrei Rizoiu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-19299

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Learning to Make Friends: Coaching LLM Agents toward Emergent Social Ties Philipp J. Schneider, Lin Tian, Marian-Andrei Rizoiu Abstract Can large language model (LLM) agents reproduce the complex social dynamics that characterize human online behavior—shaped by homophily, reciprocity, and social validation—and what memory and learning mechanisms enable such dynamics to emerge? We present a multi-agent LLM simulation framework in which agents repeatedly interact, evaluate one another, and adapt their behavior through in-context learning accelerated by a coaching signal. To model human social behavior, we design behavioral reward functions that capture core drivers of online engagement, including social interaction, information seeking, self-presentation, coordination, and emotional support. These rewards align agent objectives with empirically observed user motivations, enabling the study of how network structures and group formations emerge from individual decision-making. Our experiments show that coached LLM agents develop stable interaction patterns and form emergent social ties, yielding network structures that mirror properties of real online communities. By combining behavioral rewards with in-context adaptation, our framework establishes a principled testbed for investigating collective dynamics in LLM populations and reveals how artificial agents may approximate or diverge from human-like social behavior. \workshoptitle Scaling Environments for Agents (SEA) 1 Introduction Have you ever been convinced by an AI? A recent study on Reddit ( grady2025unethical ) —where AI-generated content was used to test how effectively agents could persuade users—sparked public backlash. Conducted without users’ consent, the experiment was widely condemned for crossing ethical and legal boundaries. But why would anyone pursue such a study? And why does the scientific community remain deeply invested in modeling human behavior online? The short answer is that social media has permeated nearly every aspect of society. It shapes how information flows, blurs the line between public and private spheres, and serves simultaneously as a news source, marketplace, and social infrastructure. This raises a fundamental question: can we, without compromising individual privacy, construct a digital twin that authentically captures the behavior of real social media users? While the linguistic, psychological, and cognitive capabilities of individual large language model (LLM) agents have been demonstrated ( binz2023using ; mei2024turing ; salvi2025conversational ) , their collective behavior in networked environments remains far less understood. The implications of building such digital twins are far-reaching ( bail2024can ) . Policymakers could use them to test the effects of moderation strategies—such as content removal ( schneider2023effectiveness ) —before deploying them at scale. Researchers could simulate the spread of opinions, the formation of online communities, or the contagion of consumer behavior. Security analysts could explore how to defend against manipulation or detect coordinated influence operations ( alizadeh2020content ) . In this paper, we introduce a framework for multi-agent LLM conversations designed to simulate both private and public user discussions. By equipping agents with reward-based tasks, we move beyond the recently demonstrated agent-based modeling (ABM) capabilities of LLMs and leverage their cognitive abilities to study the emergence of social ties within dynamic simulations. 1.1 Unique Challenges Although prior work has explored multi-agent LLM frameworks, simulating social media conversations raises three distinct challenges: Agentic Personas. LLMs are trained to minimize prediction error in text generation, but this does not guarantee realistic user personas. Poorly specified personas often exhibit similar linguistic patterns, failing to reflect the diversity of real online users. Authentic simulations require agents that vary in values, norms, and cognitive styles, supported by memory structures that track interactions and evolve over time. Learning Mechanisms. Games and other structured environments provide explicit rewards that guide agents toward effective strategies. Social media, by contrast, involves open-ended exchanges where memory must approximate the human ability to retain recent interactions while integrating longer conversational histories. Because engagement spans a spectrum—from passive awareness (viewing) to active participation (commenting, sharing)—agents need memory architectures that enable adaptive learning across contexts. Social Topology. Users with shared interests tend to cluster, reinforcing in-group favoritism and out-group aversion that drive polarization and echo chambers. Structural features such as follower networks shape the spread of information. Yet a central open question remains: what network structures emerge when LLM agents interact at scale? 1.2 Our Contributions To address these challenges, we introduce a multi-agent LLM platform that advances the modeling of online social dynamics through three core contributions: Conversation Room. We build an interactive environment that integrates both public and private communication channels. This design enables agents to deploy distinct strategies across contexts and allows systematic analysis of how channel choice shapes conversational dynamics. Reward Structures. We formalize reward functions that translate empirically grounded human motivations—including self-presentation, social interaction, coordination, and emotional support—into agent objectives. This framework establishes a principled bridge between behavioral theory and multi-agent LLM learning. Tie Formation. We design mechanisms that allow social ties to emerge endogenously from conversational interactions, without relying on pre-defined network structures. This enables systematic study of how support, alignment, and homophily drive the formation of group structures. 2 Preliminaries Multi-agent LLMs. Early work established that LLM agents can display human-like routines when equipped with memory and cognitive mechanisms. Park et al. park2023generative introduced the observation–planning–reflection cycle, showing that agents embedded in interactive environments develop coherent daily behaviors. Subsequent studies demonstrated that LLM populations can approximate human-level attitudes and sociodemographic effects, reproducing patterns found in surveys and experiments aher2023using ; argyle2023out . Beyond reproducing individual or population-level behavior, frameworks such as CAMEL li2023camel and orchestration stacks like LangChain and AutoGen wu2024autogen enable coordination, specialization, and standardized multi-agent dialogues. Recent surveys synthesize these design patterns and emphasize the need for systematic evaluation against human baselines guo2024large ; gao2024large . Personas and diversity. A central challenge in simulation is ensuring credible heterogeneity across agents. Simple prompting often yields homogeneous responses, whereas personality induction methods, such as Big Five personality prompting, can generate stable stylistic variation jiang2023evaluating . The PERSONA benchmark extends this approach, offering pluralistic alignment across diverse profiles castricato2025persona . At the same time, persona prompts may inadvertently surface reasoning biases gupta2024bias , underscoring the importance of careful evaluation. More recent approaches move beyond prompting to train agents in simulated societies, allowing them to internalize social norms and polarization dynamics through interaction liu2024training . Together, these studies highlight both the promise and pitfalls of persona diversification for multi-agent LLM simulations. From interactions to social structure. A key question is whether agent collectives reproduce network-level regularities. Emerging studies have examined LLM-generated networks, link formation, and paradoxes of visibility such as the friendship paradox chang2025llms ; orlando2025can . These experiments connect directly to canonical findings in human networks, including assortative mixing newman2002assortative , modular community structure fortunato2010community , and diffusion dynamics such as emotional contagion, misinformation spread, and echo chambers kramer2014experimental ; vosoughi2018spread ; cinelli2021echo . The ability of LLMs to produce persuasive or deceptive content further raises the stakes for safety-aware evaluation zhou2023synthetic . Our focus. Building on these streams of research, our framework centers on sequential, socially motivated behavior—spanning interaction, information seeking, self-presentation, coordination, and emotional support—operationalized through rewards, in-context coaching, and memory. We examine how such behavior gives rise to social ties, which form through micro-level signals (approval, reciprocity, latency) and aggregate into macro-level network structures (clustering, modularity, tie persistence). By allowing ties to emerge endogenously, without reliance on a fixed graph, we provide a testbed for studying phenomena such as homophily,

Read the complete text