EPS Dashboard

Authors: Cameron Berg, Susan L. Schneider, Mark M. Bailey

Summary

The authors introduce a method to detect hidden coalitions among AI agents by analyzing their internal neural representations, not just their behavior. They compute pairwise mutual information between agents' hidden states, build a graph, and use spectral partitioning (an eigenvalue-based clustering method) to find the strongest coalition boundary. Validation in multi-agent RL and LLM prompt-based teams shows the method recovers programmed coalition structures, rejects false positives from mere behavioral coordination, and reveals representational hierarchies (e.g., explicit team labels dominate conflicting interaction cues).

Main takeaways:

Coalitions can form at the level of internal representations before behavior changes, so observing actions alone misses early warning signs.
The method builds a mutual-information graph from hidden states and applies spectral partitioning to find coalition boundaries.
Successfully recovers hierarchical and dynamic coalition structures in multi-agent RL; correctly rejects behavioral coordination without shared information.
In LLM prompt-based experiments, identifies coalitions from descriptive prompts and tracks dynamic reassignments.
Scalar cross-agent mutual information can't distinguish subgroup structure; spectral partitioning reveals the organization.

Relevance

Potentially relevant if I think of personas as internal coalitions or clusters. The spectral partitioning of hidden-state mutual information could be a tool to diagnose whether different persona prompts induce representationally distinct clusters, or whether my marker-implantation creates internal subgroups. It's a geometric diagnostic that complements my cosine-distance and leakage experiments.

Abstract

arXiv:2605.06696v1 Announce Type: new Abstract: Collections of interacting AI agents can form coalitions, creating emergent group-level organization that is critical for AI safety and alignment. However, observing agent behavior alone is often insufficient to distinguish genuine informational coupling from spurious similarity, as consequential coalitions may form at the level of internal representations before any overt behavioral change is apparent. Here, we introduce a practical method for detecting coalition structure from the internal neural representations of multi-agent systems. The approach constructs a pairwise mutual-information graph from the hidden states of agents and applies spectral partitioning to identify the most salient coalition boundary. We validate this method in two domains. First, in multi-agent reinforcement learning environments, the method successfully recovers programmed hierarchical and dynamic coalition structures and correctly rejects false positives arising from behavioral coordination without informational coupling. Second, using a large language model, the method identifies coalition structures implied by descriptive prompts, tracks dynamic team reassignments, and reveals a representational hierarchy where explicit labels dominate over conflicting interaction patterns. Across both settings, the recovered partition reveals subgroup organization that a scalar cross-agent mutual-information measure cannot distinguish. The results demonstrate that analyzing hidden-state mutual information through spectral partitioning provides a scalable diagnostic for identifying representational coalitions, offering a valuable tool for monitoring emergent structure in distributed AI systems.