MUESLI: Symbolic Reasoning for Language-Based Economic Games

We present a symbolic-first GLEE agent that uses explicit models of incentives, uncertainty, and opponent behavior rather than an LLM at decision time. Our submissions finished 35th of 198 overall, including 10th in bargaining, 35th in persuasion, and 62nd in negotiation. Bargaining performed strongly because much of its strategically relevant state is exposed through structured actions, while persuasion remained competitive using Bayesian signal reasoning and lightweight deterministic language processing. Negotiation was the clear weakness despite a richer symbolic opponent model. We argue that this contrast suggests a useful boundary: symbolic methods can go far when strategic information is structurally exposed, while language models may be most valuable for interpreting information that remains latent in dialogue.