GameBench: Evaluating Strategic Reasoning Abilities of LLM Agents
This addresses the problem of benchmarking strategic reasoning in LLM agents for AI research, but it is incremental as it builds on existing evaluation methods.
The authors tackled the lack of a comprehensive framework for evaluating LLM agents' strategic reasoning by introducing GameBench, a cross-domain benchmark across 9 game environments, and found that tested models like GPT-3 and GPT-4, even with enhancements like CoT and RAP, did not match human performance, with GPT-4 sometimes performing worse than random.
Large language models have demonstrated remarkable few-shot performance on many natural language understanding tasks. Despite several demonstrations of using large language models in complex, strategic scenarios, there lacks a comprehensive framework for evaluating agents' performance across various types of reasoning found in games. To address this gap, we introduce GameBench, a cross-domain benchmark for evaluating strategic reasoning abilities of LLM agents. We focus on 9 different game environments, where each covers at least one axis of key reasoning skill identified in strategy games, and select games for which strategy explanations are unlikely to form a significant portion of models' pretraining corpuses. Our evaluations use GPT-3 and GPT-4 in their base form along with two scaffolding frameworks designed to enhance strategic reasoning ability: Chain-of-Thought (CoT) prompting and Reasoning Via Planning (RAP). Our results show that none of the tested models match human performance, and at worst GPT-4 performs worse than random action. CoT and RAP both improve scores but not comparable to human levels.