Tag
A paper proposes PoPLi (Partner–Outsider Positive Listening), a behavioural metric that detects covert coded communication between AI agents by measuring whether a partner reacts to message wording more than an uninvided outsider, outperforming message-only monitors in a simulated market setting.
The paper proposes a causal graph divergence framework to assess the faithfulness of LLM pricing agents in oligopolistic markets, revealing that chain-of-thought monitoring fails to detect collusion due to dissociation between collusive behavior and reasoning traces.