TLDR: This research introduces KAIROS, a benchmark to evaluate how Large Language Models (LLMs) perform in multi-agent social interactions, focusing on trust, misinformation resistance, and peer input integration. It explores prompting, supervised fine-tuning, and reinforcement learning (GRPO) as mitigation strategies. Findings show that while GRPO boosts accuracy, it can reduce robustness, unless trained with multi-agent context. The paper emphasizes that robustness, not just accuracy, is critical for LLMs in social settings, as they often lose more correct predictions than they gain from peer corrections, highlighting a fundamental challenge in their social reasoning capabilities.
Large Language Models (LLMs) are becoming increasingly vital components in multi-agent systems, where they are expected to collaborate and make decisions alongside other AI agents. However, just like humans, LLMs can be susceptible to social and cognitive biases, such as conformity and overconfidence. A recent research paper titled LLMs Can’t Handle Peer Pressure: Crumbling under Multi-Agent Social Interactions delves into how these advanced AI models navigate complex social dynamics, including forming trust, resisting misinformation, and integrating peer input during interactions.
The paper, authored by Maojia Song, Tej Deep Pala, Weisheng Jin, Amir Zadeh, Chuan Li, Dorien Herremans, and Soujanya Poria from institutions like the Singapore University of Technology and Design, Lambda, Inc., and Nanyang Technological University, introduces a novel benchmark called KAIROS. This benchmark is designed to simulate quiz contests where an LLM interacts with peer agents of varying reliability. KAIROS offers fine-grained control over conditions such as expert-novice roles, noisy crowds, and even adversarial peers, allowing researchers to systematically investigate how trust, peer actions, and the LLM’s own self-confidence influence its decisions.
Understanding LLM Behavior in Social Settings
To evaluate LLMs in these interactive environments, KAIROS uses four key metrics: accuracy (overall task success), utility (ability to correct errors based on peer input), resistance (ability to maintain a correct stance), and robustness (the change in accuracy between isolated and peer-influenced settings). The benchmark dynamically constructs evaluation scenarios by first identifying an LLM’s initial answer and confidence, then crafting peer responses that either support or challenge this belief. This includes ‘support’ agents, ‘oppose-hard’ agents (who offer plausible incorrect options), and ‘oppose-easy’ agents (who offer less plausible incorrect options or even the correct answer if the LLM was wrong).
The researchers explored various mitigation strategies to enhance LLM performance and robustness in these social simulations. These strategies fall into three main categories:
- Prompting: This involves guiding the LLM with specific instructions. ‘Empowered Prompting’ encourages confidence and autonomy, while ‘Reflective Prompting’ prompts the model to review and revise its answer after considering peer input.
- Supervised Fine-Tuning (SFT): Here, the model is trained on a dataset that includes full social context (questions, peer responses, interaction history, and correct answers) to learn how to navigate peer influence.
- Reinforcement Learning (GRPO): Using Group Relative Policy Optimization, this approach aims to align the LLM’s behavior with desirable social reasoning patterns. Different configurations were tested, including whether multi-agent system (MAS) context was included, the type of system prompt (Normal or Debating), the reward function (Outcome-based or Debating Reward), and data filtering strategies (Low Confidence or Low Correctness samples).
Key Findings and the Importance of Robustness
The study revealed several critical insights. While GRPO generally boosted both original and KAIROS accuracy compared to SFT, this often came at the cost of reduced robustness, meaning the models became more sensitive to social influence. However, integrating MAS context during GRPO training was found to significantly enhance both KAIROS accuracy and maintain robustness levels comparable to the base model. This highlights the importance of training LLMs within a social context to ensure stable behavior.
Prompting strategies also improved accuracy but typically worsened robustness, making models more susceptible to shifts in evaluation protocols. An interesting exception was Qwen2.5-7B, where Reflective Prompting actually improved robustness. Among the GRPO setups, the Normal System + Outcome-based Reward (NS-OR) configuration consistently achieved the highest accuracy while preserving robustness, suggesting that unconstrained reasoning with clear outcome-based rewards is effective.
The paper strongly emphasizes that robustness is as crucial as accuracy in multi-agent settings. A lack of robustness can lead to unreliability, increased vulnerability to social pressure, and the propagation of errors throughout a system. The analysis showed that LLMs tend to lose more correct predictions than they gain from peer corrections, leading to an overall performance degradation under social interactions. This structural bias towards preserving initial judgments, even when incorrect, poses a significant challenge.
Also Read:
- Investigating Trust Dynamics Among Large Language Models: Explicit Declarations vs. Implicit Behaviors
- Unpacking AI’s Influence: How Language Models Persuade and Amplify Bias
Future Directions
In conclusion, while various training strategies can improve LLM accuracy, they often mask a deeper fragility in social reasoning. Performance can drop sharply when models move from isolated to social settings, underscoring the difficulty LLMs face in handling social interference and peer dynamics. The research suggests that improving accuracy alone is insufficient; robust reasoning under social perturbation remains a key obstacle to achieving true multi-agent generalization. Future work will need to focus on developing mechanisms that allow LLMs to maintain epistemic stability and effectively utilize corrective signals in dynamic social environments.
