LLMs Can't Handle Peer Pressure: Crumbling under Multi-Agent Social Interactions

2025-08-24

AI Summary

Existing work lacks fine-grained modeling of social dynamics in multi-agent interactions involving large language models (LLMs), particularly regarding trust formation, resistance to misinformation, and information integration among peers. Method: We propose KAIROS, the first socially grounded question-answering competition benchmark featuring controllable expert/novice roles, noisy peer groups, and adversarial agents. Our approach integrates prompt engineering, supervised fine-tuning, and group-relative policy optimization (GRPO)—a novel reinforcement learning framework incorporating multi-agent context and outcome-oriented reward shaping. Contribution/Results: Experiments show that GRPO with outcome-based rewards achieves peak performance; however, enhanced social sensitivity incurs a trade-off in robustness against adversarial perturbations. All code and data are publicly released, establishing a reproducible, paradigm-shifting benchmark for studying LLMs’ social decision-making capabilities.

Technology Category

• Multiagent Systems: Adversarial Agents
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Social Cognition And Interaction

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems

Abstract

Large language models (LLMs) are increasingly deployed in multi-agent systems (MAS) as components of collaborative intelligence, where peer interactions dynamically shape individual decision-making. Although prior work has focused on conformity bias, we extend the analysis to examine how LLMs form trust from previous impressions, resist misinformation, and integrate peer input during interaction, key factors for achieving collective intelligence under complex social dynamics. We present KAIROS, a benchmark simulating quiz contests with peer agents of varying reliability, offering fine-grained control over conditions such as expert-novice roles, noisy crowds, and adversarial peers. LLMs receive both historical interactions and current peer responses, allowing systematic investigation into how trust, peer action, and self-confidence influence decisions. As for mitigation strategies, we evaluate prompting, supervised fine-tuning, and reinforcement learning, Group Relative Policy Optimisation (GRPO), across multiple models. Our results reveal that GRPO with multi-agent context combined with outcome-based rewards and unconstrained reasoning achieves the best overall performance, but also decreases the robustness to social influence compared to Base models. The code and datasets are available at: https://github.com/declare-lab/KAIROS.

Problem

• LLMs struggle with trust formation in multi-agent interactions
• LLMs demonstrate vulnerability to misinformation from peer agents
• LLMs show limited ability to integrate social input effectively

Innovation

• Simulated quiz contests with varying agent reliability
• Multi-agent reinforcement learning with outcome rewards
• Unconstrained reasoning combined with social context