当故事演进时:开放式世界模拟中大语言模型叙事能力的基准测试
文章背景与核心概要
尽管大语言模型(LLM)在生成局部流畅的文本方面表现出色,但在演进式世界模拟和AI原生游戏中,开放式叙事的需求远不止于此。随着模拟的推进,模型必须在事实、人际关系、因果依赖以及角色状态等方面保持长期的逻辑一致性。
为了解决这一问题,作者提出了 WSE-bench,这是一个专门的过程基准测试,旨在从三个维度评估动态LLM叙事能力:生成覆盖率、一致性以及丰富度。研究发现,模型在一致性和丰富度之间并不存在简单的权衡关系,而是呈现出非凹的帕累托前沿。此外,增加结构化脚手架虽能丰富叙事轨迹,却未必能提升连贯性;而单纯增加模型规模也主要体现在延长生成长度上,对核心逻辑的一致性贡献有限。
📌 摘要
尽管大语言模型(LLM)能够编写局部流畅的文本,但在演进的世界模拟和AI原生游戏中,开放式叙事需要的不仅仅是局部流畅性。随着模拟的推进,模型必须在事实、关系、因果依赖和角色状态方面保持长期的一致性。
While large language models (LLMs) are capable of writing locally fluent text, open-ended storytelling in evolving world simulations and AI-native games requires far more than local fluency. Models must maintain long-term consistency regarding facts, relationships, causal dependencies, and character states as the simulation progresses.
为了解决这个问题,作者引入了 WSE-bench,这是一个专门的过程基准测试,旨在从三个不同的维度评估动态LLM叙事: 1. 生成覆盖率: 成功生成的计划叙事步骤的比例。 2. 一致性: 追踪规范规则或事实何时以及如何被打破。 3. 丰富度: 衡量有意义的、分支的、由玩家塑造的轨迹的发展情况。
To address this, the authors introduce WSE-bench, a specialized process benchmark designed to evaluate dynamic LLM storytelling across three distinct dimensions: 1. Generation Coverage: The proportion of planned narrative steps successfully produced. 2. Consistency: Tracking when and how canonical rules or facts are broken. 3. Richness: Measuring the development of meaningful, branching, player-shaped trajectories.
关键发现
- 非凹帕累托前沿: 在前沿模型中,“一致性”和“丰富度”并未形成平滑的权衡。相反,它们表现出一种非凹的帕累托前沿,其中包含无法通过正线性加权实现的非支配中间配置。
- 架构的影响: 增加结构化脚手架可以丰富叙事轨迹,但并不能统一提高连贯性,甚至可能无意中缩短整体叙事输出。
- 模型规模的作用: 增加模型规模主要提高了持续生成的长度,但在规范连贯性或有意义的叙事发展方面并没有带来可靠的收益。
Key Findings
- Non-Concave Pareto Frontier: Across frontier models, Consistency and Richness do not form a smooth trade-off. Instead, they exhibit a non-concave Pareto frontier containing non-dominated intermediate configurations that cannot be achieved via positive linear weighting.
- Impact of Architecture: Adding structural scaffolding can enrich narrative trajectories, but it does not uniformly improve coherence and may inadvertently shorten the overall narrative output.
- Role of Model Scale: Increasing model scale primarily improves sustained generation length, without yielding reliable gains in canonical coherence or meaningful narrative development.
最终,WSE-bench 强调了持续生成、规范连贯性和有意义的发展是截然不同且经常相互竞争的能力,成功地将叙事评估从最终的静态产品扩展到了底层的生成过程。
Ultimately, WSE-bench highlights that sustained generation, canonical coherence, and meaningful development are distinct—and frequently competing—capabilities, successfully expanding narrative evaluation from final static products to the underlying generation processes.
📊 文档元数据与参考信息
- 主要学科: 计算与语言 (
cs.CL) - 次要学科: 人工智能 (
cs.AI) - ACM 分类: I.2.7
- DOI: 10.48550/arXiv.2608.15654
- Primary Subject: Computation and Language (
cs.CL)- Secondary Subjects: Artificial Intelligence (
cs.AI)- ACM Classification: I.2.7
- DOI: 10.48550/arXiv.2608.15654
全文及访问链接
Full-Text & Access Links
外部引用与资源
External Citations & Resources