arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.14570cs.AIcs.SI
收藏

PsyMemBehav:面向社交媒体行为模拟的、基于心理学与记忆的两级身份-动作推理方法

PsyMemBehav: Psychology-Informed and Memory-Grounded Two-Level Identity-to-Action Reasoning for Social-Media Behavior Simulation

Communication University of China(中国传媒大学)

Wenji Zhang, Xilong Cheng, Yunxiao Qin, Mai Tu, Tongye Zhu, Yuhao Zhou, JingBo Ji, Yuting Tan, Duoyin Li, Hongjiang Xiao, Yuan Zhang

AI总结:

本研究提出PsyMemBehav模型,构建了含5862名用户557万条行为的PsyMemBehav-bench数据集,该模型在12项个体指标中排第1,提升了回复质量并降低了群体层面误差,可用于社交媒体行为模拟。

摘要:

大型语言模型(LLM)可支持社会模拟,但要忠实复现真实用户,需将结构化平台决策与语言锚定到稀疏、含噪的行为证据中。本文提出PsyMemBehav,该方法通过两级身份-动作推理,将心理学驱动的用户画像与检索到的记忆映射为以目标为导向的平台行为。为给该方法提供全面的行为轨迹支撑,我们构建了数据集PsyMemBehav-bench,其来自Reddit、Bluesky和Twitter平台上5862名真实用户的557万条观测行为(约每名用户950条)。每条记录的互动被建模为目标选择,随后执行评论、点赞或转发操作,还可选择生成回复。PsyMemBehav结合了41维用户描述符、Top-5 BM25记忆、反向思维链(reverse-CoT)监督,以及基于 rubric( rubric 保留英文原名)的完整推理-动作轨迹的直接偏好优化(DPO)。实验表明,270亿参数规模的PsyMemBehav在12项个体层面指标中排名第1,将平均回复质量从GPT-5.4的54.28提升至59.68(相对提升9.95%),且在所有模拟系统中获得最低的群体层面平均绝对误差(MAE,0.081)与均方误差(MSE,0.009)。这些结果为观测线程延续的证据锚定用户建模提供了支撑;该基准依赖于记录的互动,不模拟弃权(不执行)决策。

英文摘要:

Large language models (LLMs) support social simulation, but faithfully reproducing real users requires grounding both structured platform decisions and language in sparse, noisy behavioral evidence. We introduce PsyMemBehav, which maps psychology-informed profiles and retrieved memories to target-first platform behavior through two-level identity-to-action reasoning. To support this approach with comprehensive behavioral traces, we construct a dataset PsyMemBehav-bench from 5.57M observed behaviors of 5,862 real users (approximately 950 per user) across Reddit, Bluesky, and Twitter. Each recorded engagement is modeled as target selection followed by Comment, Like, or Repost and optional response generation. PsyMemBehav combines 41-dimensional user descriptors, top-5 BM25 memories, reverse-CoT supervision, and rubric-guided DPO over complete reasoning-action trajectories. Experiments show that PsyMemBehav 27B ranks first in 11 of 12 reported individual-level metrics, improving average response quality over GPT-5.4 from 54.28 to 59.68 (9.95% relative), and obtains the lowest group-level MAE (0.081) and MSE (0.009) among simulated systems. These results support evidence-grounded user modeling for observed-thread continuation; the benchmark remains conditional on recorded engagement and does not model no-action decisions.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑