Research Papers 论文研究 4h ago Updated 2h ago 更新于 2小时前 49

The Story Shapes the Agent: Narrative Priors in LLM Behavior 故事塑造智能体:LLM行为中的叙事先验

Narrative framing significantly influences LLM agent behavior, explaining 5-31 times more variance than explicit persona assignments. The study isolates "narrative priors" using structurally isomorphic tasks (disease investigation, IT troubleshooting, murder mystery) to demonstrate that story context overrides decision structure. Cross-narrative persona transfer is driven by "behavioral anchors"—specific language mapping to actions—rather than abstract role descriptions. Removing behavioral anch 研究发现任务叙事的“故事框架”比角色设定更能显著影响大语言模型的行为模式,解释了5-31倍于角色的行为方差。 通过构建结构同构但叙事不同(疾病调查、IT故障排查、谋杀悬疑)的实验,隔离出“叙事先验”这一系统性行为倾向。 跨叙事有效的角色效应源于“行为锚点”,即描述中直接映射具体动作的词汇;移除这些锚点导致跨叙事一致性下降95%。 实验表明,在两个领域中叙事先验与任务成功率呈负相关,提示过度沉浸于故事框架可能损害实际表现。 研究提出了一种基于行为锚点的角色选择方法,并验证了其在未见叙事中的泛化能力,强调LLM行为应扎根于具体动作而非抽象描述。

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Narrative framing significantly influences LLM agent behavior, explaining 5-31 times more variance than explicit persona assignments.
  • The study isolates "narrative priors" using structurally isomorphic tasks (disease investigation, IT troubleshooting, murder mystery) to demonstrate that story context overrides decision structure.
  • Cross-narrative persona transfer is driven by "behavioral anchors"—specific language mapping to actions—rather than abstract role descriptions.
  • Removing behavioral anchors from persona prompts reduces cross-narrative consistency by 95%, confirming their critical role in stable agent behavior.
  • LLM behaviors intended to persist across different contexts must be grounded in concrete, action-oriented language rather than abstract character traits.

Why It Matters

This research challenges the common assumption that persona prompting is primarily about defining a character's identity, revealing instead that the narrative context acts as a stronger driver of decision-making patterns. For AI practitioners, it highlights a significant risk in deploying agents across varied domains without accounting for how story framing alters performance, potentially leading to unexpected failures in production environments. Understanding these dynamics allows for more robust prompt engineering strategies that ensure consistent agent behavior regardless of the specific scenario being addressed.

Technical Details

  • Experimental Design: Utilized structural isomorphism to create three distinct text-based investigation games (disease investigation, IT troubleshooting, murder mystery) that shared identical action spaces, stage progressions, and resource constraints, varying only the narrative framing.
  • Scale and Scope: Analyzed 1,890 sessions across three different LLM architectures and ten distinct personas to ensure findings were model-agnostic and statistically significant.
  • Key Metric: Identified "narrative priors" as systematic action tendencies activated by story framing, quantifying their impact on behavioral variance compared to persona effects.
  • Causal Intervention: Conducted experiments removing "behavioral anchors" (language mapping directly to shared actions) from high-transfer personas, resulting in a 95% drop in cross-narrative consistency.
  • Generalization: Validated the framework on a held-out fourth narrative and developed a persona-selection method that leverages behavioral anchors to improve transferability.

Industry Insight

  • Prompt Engineering Strategy: Shift focus from abstract persona definitions (e.g., "You are a helpful assistant") to concrete, action-oriented instructions that specify how to perform tasks, ensuring stability across diverse use cases.
  • Risk Mitigation: When deploying agents in multiple domains, anticipate that changing the narrative context may drastically alter behavior; implement rigorous testing across varied story framings to detect unintended shifts in decision-making.
  • Model Selection and Evaluation: Prioritize evaluation metrics that measure cross-narrative consistency rather than single-context performance, as this better reflects real-world robustness where task framing may vary.

TL;DR

  • 研究发现任务叙事的“故事框架”比角色设定更能显著影响大语言模型的行为模式,解释了5-31倍于角色的行为方差。
  • 通过构建结构同构但叙事不同(疾病调查、IT故障排查、谋杀悬疑)的实验,隔离出“叙事先验”这一系统性行为倾向。
  • 跨叙事有效的角色效应源于“行为锚点”,即描述中直接映射具体动作的词汇;移除这些锚点导致跨叙事一致性下降95%。
  • 实验表明,在两个领域中叙事先验与任务成功率呈负相关,提示过度沉浸于故事框架可能损害实际表现。
  • 研究提出了一种基于行为锚点的角色选择方法,并验证了其在未见叙事中的泛化能力,强调LLM行为应扎根于具体动作而非抽象描述。

为什么值得看

这篇文章揭示了当前LLM Agent开发中被忽视的关键变量:叙事框架对模型行为的深层操控力远超传统的角色Prompting。对于AI从业者而言,理解“叙事先验”有助于优化Agent的稳定性设计,避免模型因过度代入故事背景而偏离逻辑目标,从而提升复杂任务下的鲁棒性和成功率。

技术解析

  • 实验设计:采用结构同构法,构建三个文本调查游戏(疾病调查、IT故障排查、谋杀悬疑),保持动作空间、阶段和资源约束一致,仅改变叙事背景,以隔离叙事效应。
  • 数据规模:覆盖3种不同架构的大语言模型,10种不同角色设定,累计进行1,890次交互会话,确保结论的统计显著性和跨模型通用性。
  • 核心发现:识别出“叙事先验”(Narrative Priors),即由故事框架激活的系统性行为倾向。数据显示,叙事对行为方差的解释力是角色设定的5-31倍,且在多数情况下与任务成功负相关。
  • 机制验证:通过因果干预实验证明,角色跨叙事转移的有效性依赖于“行为锚点”(Behavioral Anchors)。移除这些指向具体行动的词汇后,角色的一致性下降了95%,证实了具体动作指令的重要性。

行业启示

  • Prompt工程优化:在设计Agent角色时,应减少抽象的性格描述,增加具体的、可执行的动作指令(行为锚点),以确保模型在不同场景下保持一致的逻辑行为。
  • 叙事风险控制:开发者需警惕“叙事先验”带来的负面偏差,特别是在需要严格遵循逻辑或流程的任务中,应避免让模型过度沉浸于故事氛围,必要时引入结构化约束来抵消叙事干扰。
  • 评估体系升级:现有的LLM评估多关注结果正确性,未来需纳入对“叙事敏感性”的评估,测试模型在相同逻辑结构但不同背景下的行为稳定性,以衡量Agent的真实泛化能力。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Agent Agent Research 科学研究 Alignment 对齐