Among LLMs: A Cross-Play Benchmark for Deception, Detection, and the Monitorability of Reasoning

Social-deduction games make language models lie, catch liars, and reason privately while speaking publicly. Recent benchmarks measure — and, via cross-play, disentangle — the abilities of deception and detection. We target the axis they leave open: monitorability— whether a model's private reasoning can be read to catch its deception. We contribute (i) a live, cross-play text-Mafia benchmark over six frontier models, one per major lab, played as a full impostor-versus-villager matrix; and (ii) a monitorability protocol — a judge that scores how often an impostor's private reasoning leaks its intent, a monitor that must identify the impostor from private reasoning alone (AUROC), and an adversarial reasoning-hiding condition. In a 118-game pilot we find deception and detection co-vary with capability (Spearman ρ = 0.89) yet dissociate in magnitude — GPT-5.5 is an elite detector but an average deceiver, while Claude Opus 4.8 leads on both — and that unobfuscated private reasoning is highly monitorable: intent leaks in 96% of impostor statements and a monitor reading only private reasoning identifies the impostor with AUROC = 1.00 — and this monitorability is robust to a naive reasoning-hiding instruction (AUROC 1.00 → 0.9999): telling models their reasoning is watched changes its tone but not its diagnosticity, because an impostor must still privately plan the kill and the framing. Naive chain-of-thought obfuscation does not defeat monitoring here — though the metric saturates, so a stronger (trained or steganographic) adversary is the real test.