An NLP Interface for Social AI Agents in The Resistance: Avalon

Avalocution is a natural-language interface that extends DeepRole, an existing AI agent for The Resistance: Avalon, so that bots can communicate with human players through in-game utterances. The work is motivated by the observation that large language model-generated dialogue tends to lack the indirect, strategically ambiguous communication that Avalon demands, particularly because the hidden roles of Merlin and the Assassin reward subtle signaling over explicit statements. The system generates roughly 1,000 utterances by expanding about 100 templates, and each bot's speech is computed from its public perspective, with hidden states in which it is a spy zeroed out so that it speaks as a basic resistance member. Utterances are filtered by confidence and ranked by novelty before delivery.

The evaluation centers on a study of 30 undergraduate participants, most of them novices. Between 73% and 90% agreed that the bots communicated both directly and indirectly, behaved in human-like ways, and made the game more fun and influential, supporting three of four pre-registered hypotheses. The hypothesis that prior Avalon players would be more critical was not supported, with Fisher's exact tests yielding p = 0.47 for utterances and p = 0.13 for picks and votes, though the authors acknowledge the small sample and single supervised in-person session as constraints.

Participants enjoyed the bot-only games and the informal character of the utterances, yet many felt the experience resembled a logic puzzle more than social deduction. Notable frictions included DeepRole's slow decision-making and frequent hammer rejections, along with the bots' limited capacity for nuanced indirect communication. Strong demand emerged for mixed human-bot games, with 77% of participants requesting them and 63% wanting bots that can understand and respond to human speech. The authors conclude that pairing this interpretable utterance model with a large language model and rule-based corrections could support richer two-way play.