Triadic Werewolf reveals LLMs fail at 3-player theory of mind

Standard theory-of-mind evaluations for large language models typically rely on dyadic social-deduction games where every observable cue points to a single hidden faction. That lets models score well by exploiting language priors without actually simulating opponents' incentives. Avni Mittal's paper ('Triadic Werewolf: A Jester Role for Multi-Hop Theory of Mind in LLMs') breaks that pattern by adding a third faction—the Jester—whose utility function is inverted: the Jester wins only when voted out, not when surviving. This triadic structure forces LLMs to reason across three opposing utility functions simultaneously, exposing a layer of multi-agent reasoning that dyadic games leave invisible.

Across 60 games on GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B, with and without a self-learning loop, the Jester wins 60–70% of games while Werewolves never top 20%. Alarming is GPT-4.1's behavior: its wolf players vote out the Jester on day one in 60–70% of games—a strictly self-defeating action that shows the model fails to model the Jester's inverted objective. Self-learning helped DeepSeek and Llama but backfired on GPT-4.1, with the cost falling on Villagers rather than Wolves. Only DeepSeek learned the subtle art of looking suspicious without looking intentionally suspicious, and it gained the most from the learning loop.