UNIST method trains small LLM agents to infer hidden roles, lifting Werewolf win rate from 41.6% to 65.6%

MARBO, accepted at EMNLP 2026, rewards agents only when their actions rest on reliable beliefs about other players' roles; UNIST reports Among Us win rates rising from 36.4% to 49.0%, with results from game benchmarks rather than real-world deception detection.