AlphaWolf

AlphaWolf is a self-play pipeline that teaches LLMs social deception through the game of Werewolf. A wolf agent plays, reflects on losses, rewinds to the critical decision point, and replays with an improved strategy. This approach resulted in Qwen3-8B-Thinking going from a 0.3 win rate to a 0.7 win rate over the span of only 45 games and 9 total gradient steps. The winning and losing transcripts become Direct Preference Optimization training pairs, while real-time consensus polling (every player votes on "who's suspicious?" after each speech) provides a novel auxiliary signal trained via ZIP-RC (ICLR 2026) to bootstrap a value function to the model. The result: a recursive self-improvement loop where the wolf gets better at deception, forcing villagers to adapt, driving emergent theory-of-mind.