Triadic Werewolf: A Jester Role for Multi-Hop Theory of Mind in LLMs

2026-06-26

AI Summary

Current evaluations of theory of mind in large language models predominantly rely on binary social reasoning tasks, which struggle to disentangle linguistic priors from genuine intention modeling. This work proposes a Werewolf-based triadic adversarial framework that introduces a “Joker” role to create utility-conflicting scenarios, compelling models to perform multi-hop theory-of-mind reasoning for optimal strategic behavior. This design uniquely reveals hierarchical multi-agent reasoning capacities obscured in binary setups and leverages the Joker’s inverted win condition to probe strategic depth. Experiments across 60 games with GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B show Joker win rates of 60–70% versus Werewolves’ sub-20% success. Notably, only DeepSeek demonstrates higher-order strategies—appearing suspicious without deliberate intent—and exhibits marked performance gains under self-play learning.

Abstract

Theory-of-mind evaluations of large language models typically use dyadic social-deduction games, where every observable cue points to a single hidden side, so a model with strong language priors can score well without ever simulating opponents' incentives. We extend the Werewolf game with a Jester, a third faction whose utility on peer suspicion is inverted because it wins by being voted out, so optimal play requires reasoning across three opposing utility functions. Across 60 games on GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B with Jester self-learning on and off, the Jester wins 60-70% of games while Werewolves never exceed 20%, and GPT-4.1 wolves vote the Jester out on day 1 in 60-70% of games, a strictly self-defeating action. Self-learning helps DeepSeek and Llama but hurts GPT-4.1, with the cost landing on Villagers rather than Werewolves. Only DeepSeek learns the subtle strategy of looking suspicious without looking intentionally suspicious, and it gains the most from the loop. Triadic incentive structure exposes a layer of multi-agent reasoning that dyadic deduction games leave invisible.

Problem

• theory of mind
• social deduction
• multi-agent reasoning
• triadic incentives
• large language models

Innovation

• Triadic Werewolf
• Jester role
• multi-hop Theory of Mind
• incentive structure
• self-learning