New spectral diagnostic uncovers hidden coalitions in multi-agent AI systems

A new paper from researchers Cameron Berg, Susan L. Schneider, and Mark M. Bailey, titled 'Hidden Coalitions in Multi-Agent AI: A Spectral Diagnostic from Internal Representations,' addresses a critical blind spot in AI safety: emergent coalitions among interacting AI agents. These coalitions can form at the level of internal representations, long before any overt behavioral coordination becomes visible. To detect them, the team proposes constructing a pairwise mutual-information graph from agents' hidden states and applying spectral partitioning to reveal the most salient coalition boundaries. This approach distinguishes genuine informational coupling from spurious behavioral similarity, a key limitation of existing scalar measures.

The method was validated in two distinct domains. In multi-agent reinforcement learning environments, it successfully recovered programmed hierarchical and dynamic coalition structures and correctly rejected false positives from agents that coordinated behaviorally without sharing information. In experiments with a large language model, spectral partitioning identified coalition structures implied by descriptive prompts, tracked dynamic team reassignments, and revealed a representational hierarchy where explicit labels dominated over conflicting interaction patterns. According to the authors, this provides a scalable diagnostic for monitoring emergent structure in distributed AI systems. For AI safety researchers, the technique offers a practical tool to detect hidden alignment risks—such as agents forming covert alliances that could undermine system objectives—before they manifest in observable behavior.