Structured Opponent Modeling for LLM Agents via Causal Graphs

A new framework called Structured Opponent Modeling (SOM) enables LLM-based agents to predict opponents' behavior more accurately in multi-agent environments. SOM separates opponent model construction from prediction, using a Structural Causal Model (SCM) to capture directed links between observations and actions. The LLM then performs structured reasoning along SCM-derived pathways, improving prediction accuracy and stability. The approach addresses limitations of existing methods that entangle modeling with prediction and rely on implicit reasoning. The paper is available on arXiv under ID 2605.07301.