跳转至

文章背景与核心概要

随着具备人格特征的大语言模型智能体（PC-Agents）越来越多地被部署在需要长期连贯性的角色中（如情感支持、社会模拟和互动角色扮演），确保其行为的一致性与真实性变得至关重要。这种连贯性的基石在于个性演变（Personality Evolution）——即智能体在遭遇重大生活事件后，展现出符合心理学规律的、合理的个性转变能力。

本文以“大五人格（Big Five）”心理学框架为基础，深入探讨了LLM智能体在经历11种重大生活事件后的适应机制。通过对14种模型进行测试，作者推出了BFI-Adapt这一可重用的基准测试，旨在对事件引发的个性变化的方向保真度进行评分。研究表明，当前的PC-Agents虽然能够捕捉到人类个性动态的大致趋势，但在复制细粒度的分布和效应量（effect sizes）方面仍存在不足。

Do AI Personas Grow? Analyzing and Benchmarking Personality Evolution in LLM Agents After Life Events

arXiv: [2608.06485 [cs.CL]]
Submitted: August 6, 2026
Primary Subject: Computation and Language (cs.CL)
Other Subjects: Artificial Intelligence (cs.AI), Social and Information Networks (cs.SI)
Authors: Ming Wang, Peidong Wang, Xiaocui Yang, Daling Wang, Shi Feng, Fiona Fui-Hoon Nah, Ee-Peng Lim

arXiv: [2608.06485 [cs.CL]]
Submitted: August 6, 2026
Primary Subject: Computation and Language (cs.CL)
Other Subjects: Artificial Intelligence (cs.AI), Social and Information Networks (cs.SI)
Authors: Ming Wang, Peidong Wang, Xiaocui Yang, Daling Wang, Shi Feng, Fiona Fui-Hoon Nah, Ee-Peng Lim

📋 Executive Summary

人格条件化大语言模型智能体（PC-Agents）正日益被部署于需要长期连贯性的角色中，例如情感支持、社会模拟以及交互式角色扮演。这种连贯性的基石是个性演变——即智能体在遭遇重大生活事件后，经历合理且基于心理学的个性转变的能力。

本研究通过使用“大五人格”心理学框架，探讨了LLM智能体在11个重大生活事件后的适应情况。通过测试14个模型，作者推出了BFI-Adapt，这是一个旨在评分事件引发的个性变化方向保真度的可重用基准，并揭示了当前的PC-Agents能够捕捉人类个性动态的总体趋势，但无法复制其细粒度分布和效应量。

📋 Executive Summary

Personality-conditioned Large Language Model agents (PC-Agents) are increasingly deployed in roles requiring long-term coherence, such as emotional support, social simulation, and interactive role-playing. A cornerstone of this coherence is personality evolution—the capacity of an agent to experience plausible, psychology-grounded shifts in personality after encountering significant life events.

This paper investigates how LLM agents adapt following 11 major life events using the "Big Five" personality traits as a psychometric framework. By testing 14 models, the authors introduce BFI-Adapt, a reusable benchmark designed to score the directional fidelity of event-induced personality changes, and reveal that current PC-Agents capture the general trends of human personality dynamics while failing to replicate their fine-grained distribution and effect sizes.

🔍 Key Findings & Insights

  • 可测量的特质转变： PC-Agents在经历生活事件后会表现出清晰的特质调整，无论事件与特质对是否符合有据可查的人类轨迹，其发生频率都具有可比性。
  • 低估的效应量： 尽管转变常常朝着预期的心理学方向发展，但其总体幅度通常远低于真实人类的效应量范围。
  • 有限的人口统计调节作用： 指示性别或文化区域的标准提示词对轨迹的调节作用微乎其微。
  • 压缩的离散度： 相对于实际人类样本群体，角色层面的离散度被压缩了3到4倍。
  • 稳健的响应模式： 验证测试证实，观察到的转变超过了基线无事件重测噪声，在独立的提示词释义下保持稳定，与行为选择表现出有限且与模型相关的对齐，并且能够通过不相关的介入对话持续存在。

🔍 Key Findings & Insights

  • Measurable Trait Shifts: PC-Agents exhibit clear trait adjustments following life events, occurring at comparable rates regardless of whether the event-trait pair matches documented human trajectories.
  • Underestimated Effect Sizes: While shifts frequently move in expected psychological directions, their overall magnitudes typically fall well below real human effect-size ranges.
  • Limited Demographic Moderation: Standard prompts indicating gender or cultural regions demonstrate minimal moderating effect on the trajectories.
  • Compressed Dispersion: Persona-level dispersion is compressed by 3x to 4x relative to actual human sample populations.
  • Robust Response Patterns: Validation testing confirms that the observed shifts exceed baseline no-event retest noise, remain stable against independent prompt paraphrasing, show limited and model-dependent alignment with behavioral choices, and persist through unrelated intervening dialogues.

📊 Benchmark & Evaluation: BFI-Adapt

为了系统地评估各个模型的个性发展，本研究引入了 BFI-Adapt： * 目的： 一个强大的评分基准，用于衡量由生活事件驱动的个性演变的方向保真度。 * 范围： 应用于对 14个不同的模型 进行排名，以比较不同架构模拟类人心理演变的有效性。

📊 Benchmark & Evaluation: BFI-Adapt

To systematically evaluate personality development across models, the study introduces BFI-Adapt: * Purpose: A robust scoring benchmark measuring the directional fidelity of personality evolution driven by life events. * Scope: Applied to rank 14 distinct models to compare how effectively different architectures simulate human-like psychological evolution.

🔗 Access & Resources

🔗 Access & Resources