Large Language Models Empowered Personalized Web Agents (2410.17236v2)

Abstract: Web agents have emerged as a promising direction to automate Web task completion based on user instructions, significantly enhancing user experience. Recently, Web agents have evolved from traditional agents to LLMs-based Web agents. Despite their success, existing LLM-based Web agents overlook the importance of personalized data (e.g., user profiles and historical Web behaviors) in assisting the understanding of users' personalized instructions and executing customized actions. To overcome the limitation, we first formulate the task of LLM-empowered personalized Web agents, which integrate personalized data and user instructions to personalize instruction comprehension and action execution. To address the absence of a comprehensive evaluation benchmark, we construct a Personalized Web Agent Benchmark (PersonalWAB), featuring user instructions, personalized user data, Web functions, and two evaluation paradigms across three personalized Web tasks. Moreover, we propose a Personalized User Memory-enhanced Alignment (PUMA) framework to adapt LLMs to the personalized Web agent task. PUMA utilizes a memory bank with a task-specific retrieval strategy to filter relevant historical Web behaviors. Based on the behaviors, PUMA then aligns LLMs for personalized action execution through fine-tuning and direct preference optimization. Extensive experiments validate the superiority of PUMA over existing Web agents on PersonalWAB.

• The paper introduces the Personalized User Memory-enhanced Alignment (PUMA) framework that integrates user profiles with LLMs to improve task comprehension and execution.
• It presents the PersonalWAB benchmark, rigorously evaluating personalized search, recommendation, and review generation tasks in simulated web environments.
• Experimental results demonstrate PUMA's superior function and result accuracy with a reduced memory footprint and efficient multi-turn interactions compared to baseline models.

LLMs Empowered Personalized Web Agents: An Expert Overview

The paper "LLMs Empowered Personalized Web Agents" explores the development and deployment of web agents that integrate LLMs with personalized user data. This integrated approach addresses the limitations of traditional LLM-based web agents by leveraging user profiles and historical web behaviors to enhance the comprehension and execution of personalized user instructions.

Core Contributions and Methodology

The authors introduce the task of LLM-empowered personalized web agents, focusing on integrating personalized data to improve task comprehension and execution. The task is formalized with several key elements: user profiles, personalized instructions, and a simulated web environment represented by web functions. User profiles encompass both static attributes and dynamic behaviors, while web functions are abstracted interactions with web services to facilitate personalized outcomes.

To evaluate this task, the authors present the Personalized Web Agent Benchmark (PersonalWAB), which includes diverse user instructions and personalized data across three tasks: personalized search, recommendation, and review generation. This benchmark serves as a foundation for evaluating the personalization capabilities of web agents.

The paper proposes a Personalized User Memory-enhanced Alignment (PUMA) framework, which utilizes a memory bank for task-specific retrieval of relevant historical behaviors. PUMA adapts LLMs for personalized web tasks by fine-tuning with task-specific data and optimizing instruction execution through direct preference optimization.

Results and Evaluation

Extensive experiments demonstrate that PUMA outperforms existing baseline models, showcasing improvements in both function accuracy and result accuracy in single-turn instructions. The framework's efficiency is highlighted by its ability to execute tasks effectively with a reduced memory footprint and a smaller LLM, particularly when compared to baselines using larger models and more extensive memory.

In multi-turn interactions, PUMA continues to excel by effectively utilizing feedback loops, enabling more accurate function calls and parameter settings, thus leading to higher completion rates with fewer interactions.

Implications and Future Directions

The paper's findings highlight significant potential for personalized web agents in enhancing user interaction with web services. By effectively incorporating user profiles and behavioral data, personalized web agents can significantly improve user satisfaction through more tailored service offerings.

Future research directions may include expanding the range of application domains of personalized web agents, integrating advanced user modeling techniques to dynamically adapt to evolving user preferences, and refining interaction mechanisms for user-in-the-loop scenarios. These efforts can further optimize the balance between user autonomy and agent intervention, ultimately enriching personalized web interactions.

Conclusion

This paper marks a notable advancement in the development of LLM-based personalized web agents. By addressing key personalization challenges and proposing a robust benchmark and framework, it sets a precedent for future innovations in personalized AI agents. The contributions outlined offer a comprehensive foundation for subsequent research and development in personalized web service automation.