Towards a Strong, Human-Compatible Codenames AI Agent

Abstract

Current language models are limited in their ability to solve complex planning and reasoning problems without the aid of search procedures. While a large body of work has developed search procedures tailored to single-turn, single-user natural language interactions, language generation in multi-agent contexts involving multiple users, imperfect information, and partially misaligned objectives remains extremely challenging. We aim to build search procedures that will enable language models to assist with interactive, multi-agent decision-making in a diverse range of contexts. Using the word game Codenames as a benchmark, we will combine game-theoretic planning procedures with basic language model-based scoring methods to create agents that both play strong policies and play well with human policies. This work yields a set of practical text generation procedures, new evaluation benchmarks, and foundational algorithmic improvements in language model search.