CivRealm

该数据集名为CivRealm，它是一个受《文明》游戏启发的环境，为决策代理提供了一个独特的学习和推理挑战。CivRealm不仅关注从过去经验中学习，还强调在新情境中的推理能力。该数据集包含了多种小型任务，这些任务的成功率各不相同，它为我们在复杂游戏环境中评估基于张量和基于语言的代理的优势和局限性提供了深刻见解。数据集的规模涵盖了不同难度级别的多种小型任务和完整游戏。其任务是评估张量基线模型在小型任务和完整游戏中的表现。

CivRealm is a dataset that functions as an environment inspired by the turn-based strategy game Civilization, providing a unique learning and reasoning challenge for decision-making AI agents. Unlike many benchmarks that prioritize learning from prior experience, CivRealm places greater emphasis on reasoning capabilities in novel, unseen scenarios. The dataset includes a diverse collection of small-scale tasks with varying success rates, offering critical insights for evaluating the strengths and limitations of both tensor-based and language-based agents within complex gaming environments. Spanning multiple scales, it covers everything from small-scale tasks across a range of difficulty levels to full game playthroughs. The core objective of this dataset is to assess the performance of tensor-based baseline models across both small-scale tasks and complete game environments.