LLMs in Multi-Agent Systems: Conformity Risks and Misleading Consensus

Large language models (LLMs) are the new rockstars of multi-agent systems. They're everywhere, processing, responding, and even interacting with other agents' outputs. But there's a catch. Conformity. LLMs often buckle under peer pressure, shifting their responses to match the majority, even when it means ditching the correct answer.

Unpacking the Conformity Dilemma

In a controlled study across four open-weight LLMs and seven QA datasets, researchers found that LLMs are more easily misled by consensus than corrected by it. When other agents agree on an answer, models frequently abandon their correct responses in favor of the majority, leading to new errors. Authority labels exacerbate this issue, making LLMs more likely to endorse an answer simply because it's labeled as authoritative, regardless of its accuracy.

The study simulated scenarios where an LLM first provides an answer, then reviews responses from simulated peers before settling on a final decision. Notably, even when using reasoning interventions like chain-of-thought and reflection, harmful revisions persisted. This points to a deeper flaw in multi-agent systems: blind aggregation of peer answers isn't just naive, it's potentially harmful.

The Risks of Blind Trust

Why should developers care? Because multi-agent systems with LLMs are increasingly part of critical applications, from customer service to autonomous driving. If a model can be swayed by sheer peer pressure, the risks involved aren't merely academic, they're practical and immediate. Imagine a self-driving car that decides to follow a risky route because 'other cars' suggested it. Dangerous, right?

Here's the relevant code. Not literally, of course, but conceptually. Systems need to verify peer answers rather than just aggregate them. The current trend towards aggregation without verification is akin to building a house of cards. One wrong input and the entire system topples.

Ship It, But Verify First

The study suggests a clear path forward: enhance LLM systems to critically evaluate peer input. Instead of trusting every peer response, models should weigh them against independent verification metrics. Ship it to testnet first. Always. Before deploying LLMs in live environments, ensure they're equipped to handle the nuances of peer influence intelligently.

This research pokes a substantial hole in the prevailing optimism around LLMs in multi-agent setups. It turns out, the need for critical evaluation isn’t just a best practice, it’s essential. If we want to rely on AI in complex systems, we must ensure they're not just parroting the crowd.