PIANIST: Learning Partially Observable World Models with LLMs for Multi-Agent Decision Making

PIANIST framework decomposes complex decision-making tasks into seven components for LLMs, enabling zero-shot generation of a working world model for multi-agent, partial information scenarios. It uses game descriptions and observation formats to simulate MCTS, demonstrating effectiveness in language and non-language games without domain-specific training or explicit world model definitions. PIANIST combines state prediction, reward assignment, and information partitioning for multi-player games, using an information realization function to map information sets to hidden states. It integrates with Monte Carlo Tree Search (MCTS) by sampling hidden states and simulating actions with highest UCT values, averaging across states within the same information set. PIANIST, using an LLM-generated world model, performed comparably to ground-truth models in games, indicating accurate world model creation. It outperformed direct LLM action queries, showing enhanced planning. However, PIANIST struggled against humans in GOPS and Taboo, highlighting the need for improved LLM decision-making in complex environments. Despite this, PIANIST demonstrated strong performance, suggesting potential for more nuanced adaptations to balance decision-making across different games, enhancing robust generalization in multi-agent settings.