🤖 AI Summary
To address the challenge of individual credibility inference in incomplete-information multi-agent games, this paper proposes a zero-shot graph-augmented retrieval framework. It dynamically constructs trust graphs, performs multi-hop evidence-chain retrieval, and leverages large language model (LLM) prompting—without fine-tuning—to enable interpretable and robust intent analysis. The core contribution is the first zero-shot, graph-structured evidence retrieval mechanism that tightly couples trust modeling with multi-source evidence-chain reasoning, substantially enhancing interpretability and hallucination resistance. Experiments demonstrate a 50.5% improvement in inference accuracy on the *Werewolf* game benchmark and a 30.6% reduction in hallucination rate. On a real-world U.S. election Twitter dataset, accuracy increases by 10.4%, validating strong cross-domain generalizability and practical applicability.
📝 Abstract
Trustworthiness reasoning aims to enable agents in multiplayer games with incomplete information to identify potential allies and adversaries, thereby enhancing decision-making. In this paper, we introduce the graph retrieval-augmented trustworthiness reasoning (GRATR) framework, which retrieves observable evidence from the game environment to inform decision-making by large language models (LLMs) without requiring additional training, making it a zero-shot approach. Within the GRATR framework, agents first observe the actions of other players and evaluate the resulting shifts in inter-player trust, constructing a corresponding trustworthiness graph. During decision-making, the agent performs multi-hop retrieval to evaluate trustworthiness toward a specific target, where evidence chains are retrieved from multiple trusted sources to form a comprehensive assessment. Experiments in the multiplayer game emph{Werewolf} demonstrate that GRATR outperforms the alternatives, improving reasoning accuracy by 50.5% and reducing hallucination by 30.6% compared to the baseline method. Additionally, when tested on a dataset of Twitter tweets during the U.S. election period, GRATR surpasses the baseline method by 10.4% in accuracy, highlighting its potential in real-world applications such as intent analysis.