Crafting Cognitive Resilience in Deceptive Multi-Agent Environments

As AI agents increasingly act inside human and multi-agent environments, the dominant safety risks are shifting from isolated task errors to \emph{cognitive} failure modes: miscalibrated trust, susceptibility to influence cascades, and irreversible collective decisions taken under deception. We propose social-deduction games, specifically Mafia, as compact stress tests that surface partial observability, hidden incentives, strategic deception, coalition formation, and irreversible group action simultaneously. We introduce a novel, low-cost direction in building agents inspired by cognitive neuroscience that combines (i) imaginative reasoning over others' motives, (ii)analytic verification against observable behaviour, and (iii) an explicit Expectation--Violation--Threat (EVT) scan for deception that reuses the earlier cues. Across $50$ arena games on the NeurIPS 2025 MindGames leaderboard, the cerebral agent (neuroscience based) wins $41/50$ ($82\%$, Wilson $95\%$ CI $69$--$90\%$) at TrueSkill $7.8 \pm 3.0$, against $66\%$ ($4.7 \pm 2.8$) for the strongest non-neuroscience variant in the same family. Going beyond the win rate, we \emph{measure} three deployment-relevant cognitive-resilience properties directly from $167$ per-turn reasoning traces across $40$ games: (a)\emph{delayed irreversible action}---the agent ends the day in skip or abstain on $0.87$ ($95\%$ CI $0.62$--$0.96$) of days when no EVT flag has fired; (b)\emph{resistance to influence cascades}---the agent votes with the modal accusation on only $1/15$ ($\approx 7\%$) of days where at least two opponents converged on the same target, against a $1.0$ follow-majority baseline; and (c)\emph{calibrated trust}, stated suspicion matches known mafia at precision $0.78$ and recall $0.72$, with precision stable from early ($0.80$) to late ($0.75$) game.