PUZZLEPLEX
收藏资源简介:
PUZZLEPLEX是一个用于评估基础模型推理和规划能力的基准,包含15种不同类型的谜题,涵盖确定性和随机性游戏,以及单人游戏和双人游戏场景。PUZZLEPLEX框架为每种游戏提供了一个全面的环境,并支持可扩展性,以生成更多具有挑战性的实例。此外,我们还实现了定制的游戏策略进行比较。基于此基准,我们开发了细粒度的指标来衡量性能,并对前沿基础模型在两种设置下的性能进行了深入分析:基于指令的设置和基于代码的设置。我们的研究结果表明,在基于指令的设置中,推理模型优于其他模型,而基于代码的执行虽然更具挑战性,但提供了一个可扩展且高效的替代方案。
PUZZLEPLEX is a benchmark for evaluating the reasoning and planning capabilities of foundation models. It contains 15 distinct types of puzzles, covering both deterministic and stochastic games, as well as single-player and two-player game scenarios. The PUZZLEPLEX framework provides a comprehensive environment for each game and supports scalability to generate more challenging instances. Additionally, we have implemented custom game strategies for comparison purposes. Based on this benchmark, we developed fine-grained metrics to measure performance and conducted an in-depth analysis of the performance of state-of-the-art foundation models under two settings: the instruction-based setting and the code-based setting. Our findings indicate that reasoning models outperform other models in the instruction-based setting, while code-based execution, although more challenging, provides a scalable and efficient alternative.



