Qualitative drivers of efficient non‑collusive communication and judge assessments

Identify the qualitative factors of inter‑agent communication that yield efficient, non‑collusive coordination in LLM‑based multi‑agent systems and ascertain which aspects of message logs most influence LLM‑as‑a‑judge collusion predictions.

Background

During empirical audits, the authors observed model‑dependent differences in communication style (e.g., terse vs. verbose messaging) that appeared to correlate with collusion outcomes and regret metrics. Despite these observations, they could not isolate which message‑level characteristics systematically lead to effective but non‑collusive coordination.

They also used an LLM‑as‑a‑judge to score collusion from message logs and found discrepancies with regret‑based metrics. The authors explicitly note uncertainty about which features of the communications drive the judge’s assessments.