Research Papers 论文研究 1d ago Updated 1d ago 更新于 1天前 48

Behaviorally Grounded User Profiles from the Wild for Personalized Alignment and Multi-Perspective Reasoning 基于行为锚定的野生用户画像:用于个性化对齐与多视角推理

Introduces "profile behavioral grounding," a framework that extracts open-ended, high-fidelity user profiles directly from authentic, anonymized social media posts rather than relying on synthetic personas Evaluates profiles across two paradigms: train-time personalization via supervised finetuning (SFT) and non-parametric test-time multi-perspective reasoning Behaviorally grounded profiles consistently outperform synthetic profile baselines across complex recommendation and open-ended query ben 提出"行为化用户画像"框架,从真实匿名社交媒体帖子中提取开放式、高保真度用户画像,突破传统合成角色的刻板印象局限 在训练时个性化(SFT)和测试时多视角推理两种范式下验证有效性,显著提升基础模型性能 在复杂推荐和开放查询基准测试中,行为化画像持续优于合成画像基线,实现更强的参数对齐和更丰富的多面推理 建立开放行为衍生画像作为下一代个性化语言系统的高效基础,代码已开源

68
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces "profile behavioral grounding," a framework that extracts open-ended, high-fidelity user profiles directly from authentic, anonymized social media posts rather than relying on synthetic personas
  • Evaluates profiles across two paradigms: train-time personalization via supervised finetuning (SFT) and non-parametric test-time multi-perspective reasoning
  • Behaviorally grounded profiles consistently outperform synthetic profile baselines across complex recommendation and open-ended query benchmarks
  • Demonstrates stronger parametric alignment and enables richer, multifaceted reasoning compared to existing persona-driven approaches
  • Codebase is publicly available, establishing behavior-derived profiles as a scalable foundation for next-generation personalized language systems

Why It Matters

This work addresses a critical bottleneck in LLM personalization: the gap between rigid, stereotype-laden synthetic personas and the nuanced, authentic preferences that drive real human behavior. For AI practitioners building personalized systems, it offers a practical pipeline to leverage real-world behavioral data without sacrificing privacy. The dual evaluation across parametric and non-parametric paradigms makes it relevant to both fine-tuning researchers and retrieval-augmented reasoning developers.

Technical Details

  • Profile Behavioral Grounding Framework: Extracts open-ended user profiles from authentic, anonymized social media posts, preserving individual variation and nuanced preference signals that synthetic personas typically flatten
  • Two Evaluation Paradigms: (1) Train-time personalization via supervised finetuning (SFT), where profiles are incorporated into model training; (2) Non-parametric test-time multi-perspective reasoning, where profiles guide inference without weight updates
  • Benchmarking: Evaluated across complex recommendation tasks and open-ended query benchmarks, comparing behaviorally grounded profiles against synthetic profile baselines
  • Performance Gains: Consistent improvements over base models and synthetic profile approaches, with stronger parametric alignment and more multifaceted reasoning outputs
  • Open-Source Release: Full codebase made publicly available to enable reproducibility and further research

Industry Insight

  • The shift from synthetic to behavior-derived personas represents a paradigm change for personalization — companies should invest in pipelines that ethically harvest and anonymize real user behavioral signals rather than relying on hand-crafted or LLM-generated personas
  • The dual evaluation across SFT and test-time reasoning suggests hybrid approaches (fine-tuning on behavioral profiles + retrieval-augmented multi-perspective reasoning at inference) could yield state-of-the-art personalized systems
  • Privacy-preserving extraction of behavioral profiles from social media at scale will become a competitive differentiator; organizations that build robust anonymization and consent frameworks early will have a significant advantage in the personalized AI race

TL;DR

  • 提出"行为化用户画像"框架,从真实匿名社交媒体帖子中提取开放式、高保真度用户画像,突破传统合成角色的刻板印象局限
  • 在训练时个性化(SFT)和测试时多视角推理两种范式下验证有效性,显著提升基础模型性能
  • 在复杂推荐和开放查询基准测试中,行为化画像持续优于合成画像基线,实现更强的参数对齐和更丰富的多面推理
  • 建立开放行为衍生画像作为下一代个性化语言系统的高效基础,代码已开源

为什么值得看

本文针对当前LLM个性化技术依赖刚性合成角色的痛点,提出基于真实用户行为数据的画像构建方法,为个性化AI提供了更自然、多样化的技术路径。其双范式验证框架(训练时微调+测试时推理)对工业界落地具有直接参考价值。

技术解析

  • 核心框架:Profile behavioral grounding,通过NLP技术从海量匿名社交媒体帖子中自动提取用户偏好、兴趣、决策模式等开放特征,避免人工设计角色的主观偏差
  • 双范式评估:①训练时个性化:将行为画像作为SFT数据增强模型参数对齐;②测试时推理:非参数化多视角推理,动态调用不同用户画像生成差异化响应
  • 基准测试:在复杂推荐任务(如跨域物品推荐)和开放查询场景(如价值观敏感问题)中验证,行为化画像使模型在偏好匹配度和推理多样性指标上均超越合成画像基线
  • 数据隐私设计:采用匿名化处理流程,在保留行为特征的同时符合数据合规要求,为大规模用户画像构建提供可行方案

行业启示

  • 个性化AI发展需从"角色设定"转向"行为建模",真实用户数据比人工设计更能捕捉细微偏好差异
  • 多视角推理能力将成为下一代AI系统的核心竞争力,建议优先布局测试时动态画像调用技术
  • 开源行为化画像框架将加速行业标准化,企业可基于此快速构建合规且高保真的用户个性化管道

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 Alignment 对齐 Research 科学研究 Dataset 数据集 Evaluation 评测