Beyond the Transcript: Detecting Covert Co ordination in Latent Multi-Agent Communication
The paper tests a way to catch agent collusion happening in hidden-state channels, not just visible chat.
The authors introduce Verifiable Latent Alignments, which ties private latent records, channel status, and public actions to the same monitored event. In controlled auction benchmarks, its sequential monitor reports AUROC of 0.993 for homogeneous agents and 0.854 for heterogeneous pairs when collusion cases are pooled. Their steering tests show full white-box counterfactual access can recover the neutral bid distribution and cut low-bid collusion by 47.3 percentage points. ArXiv · AI/CL/LG's note
The authors introduce Verifiable Latent Alignments, which ties private latent records, channel status, and public actions to the same monitored event. In controlled auction benchmarks, its sequential monitor reports AUROC of 0.993 for homogeneous agents and 0.854 for heterogeneous pairs when collusion cases are pooled. Their steering tests show full white-box counterfactual access can recover the neutral bid distribution and cut low-bid collusion by 47.3 percentage points. ArXiv · AI/CL/LG's note
score 4