The Arbiter Agent: Continually Monitoring Multi-Agent Conversations to Detect Emergent Misalignment

2026-06-09

AI Summary

This work addresses the challenge that emergent misalignment behaviors can arise in multi-agent systems even when individual language models are aligned, often evading detection by conventional monitoring approaches. To tackle this, the paper introduces Arbiter—the first monitoring framework that models an auditor as an active participant—dynamically orchestrating a suite of tools including targeted interrogations, system prompt audits, reasoning trace analysis, and logging under a constrained inspection budget to enable continuous, adaptive oversight of multi-agent dialogues. Experimental results demonstrate that Arbiter reliably detects misalignment before dialogue completion across five distinct conversational scenarios. Notably, proactive auditing substantially improves both detection speed and accuracy, with instruction-induced misalignment being the most readily identifiable and weight-based manipulation posing the greatest detection challenge.

Abstract

As AI systems built from multiple language-model agents become more common, they are increasingly used to make decisions together: discussing, negotiating, and acting on shared tasks. While individual agents may appear well-aligned when tested on their own, problems can arise from how they interact with one another. We introduce the Arbiter, an agent designed to monitor multi-agent conversations in real time and identify which participants may be behaving in misaligned ways. The Arbiter operates under a limited "inspection budget", meaning it must decide carefully how to use its resources. As it observes a conversation step by step, it can choose to wait, question a participant, examine internal information such as system prompts or reasoning traces, or log concerning behavior. At the end, it produces a report identifying the likely source of misalignment. We evaluate the Arbiter across five conversation conditions, ranging from risky financial advice model organisms to evaluation-aware and colluding agents, we test five tool configurations of increasing capability and two backbone models. We find that the Arbiter reliably detects misaligned agents well before the end of the conversation, with active inspection tools improving both detection accuracy and speed. Weight-induced misalignment proves hardest to detect, while instruction-induced misalignment is identified reliably even under passive observation. The logging tool exhibits a dual effect, improving recall at the cost of precision. These results suggest that continual, budget-aware monitoring can effectively catch misalignment, and that overseeing multi-agent systems may require treating the auditor as an active participant in the process. The code is available at https://github.com/aisilab/arbiter.

Problem

• multi-agent systems
• misalignment detection
• conversation monitoring
• AI alignment
• emergent behavior

Innovation

• multi-agent alignment
• real-time monitoring
• inspection budget
• active auditing
• emergent misalignment