PersonaAgent: When Large Language Model Agents Meet Personalization at Test Time (2506.06254v1)

Abstract: LLM empowered agents have recently emerged as advanced paradigms that exhibit impressive capabilities in a wide range of domains and tasks. Despite their potential, current LLM agents often adopt a one-size-fits-all approach, lacking the flexibility to respond to users' varying needs and preferences. This limitation motivates us to develop PersonaAgent, the first personalized LLM agent framework designed to address versatile personalization tasks. Specifically, PersonaAgent integrates two complementary components - a personalized memory module that includes episodic and semantic memory mechanisms; a personalized action module that enables the agent to perform tool actions tailored to the user. At the core, the persona (defined as unique system prompt for each user) functions as an intermediary: it leverages insights from personalized memory to control agent actions, while the outcomes of these actions in turn refine the memory. Based on the framework, we propose a test-time user-preference alignment strategy that simulate the latest n interactions to optimize the persona prompt, ensuring real-time user preference alignment through textual loss feedback between simulated and ground-truth responses. Experimental evaluations demonstrate that PersonaAgent significantly outperforms other baseline methods by not only personalizing the action space effectively but also scaling during test-time real-world applications. These results underscore the feasibility and potential of our approach in delivering tailored, dynamic user experiences.

• The paper introduces PersonaAgent, a framework that integrates personalized memory and tailored actions to align LLM agents with individual user preferences.
• It employs dual memory modules, including episodic and semantic memories, alongside an iterative test-time alignment strategy for dynamic personalization.
• Empirical evaluations across four tasks demonstrate enhanced accuracy, improved F1 scores, and reduced errors compared to baseline models.

PersonaAgent: Personalization in LLM Agents

The paper "PersonaAgent: When LLM Agents Meet Personalization at Test Time" presents a novel framework that aims to enhance personalization within LLM agents, allowing them to better cater to individual user preferences. Traditional LLM agents often suffer from a one-size-fits-all approach, and this research seeks to address that limitation through the development of PersonaAgent—a personalized agent framework explicitly designed to tackle versatile personalization tasks.

The central innovation presented in this paper is the integration of personalization within LLM agents through two pivotal components: personalized memory and personalized actions. The personalized memory module comprises both episodic and semantic memory mechanisms. The episodic memory retains detailed, context-dependent user interactions, while semantic memory provides a stable representation of user profiles, capturing generalized traits and preferences over time. Together, these mechanisms enable the agent to maintain a coherent understanding of the user that persists across sessions.

The personalized action module empowers agents to perform tailored actions using tools relevant to the user's context and memory data. This approach significantly enhances the adaptability of LLM agents compared to general tools or policies used by traditional agents. The persona—formulated as a unique system prompt for each user—acts as an intermediary that continuously evolves through interactions. It leverages insights from memory to modulate agent actions, and those actions subsequently refine the memory, creating a robust loop for personalization.

An innovative test-time user-preference alignment strategy is introduced to further optimize personalization. By simulating recent user interactions, the framework adjusts the persona prompt based on textual discrepancies between the agent's simulated responses and the user's true preferences. This iterative optimization process utilizes textual loss feedback, ensuring that the persona remains accurately aligned with the user's evolving preferences, thus delivering dynamic and tailored user experiences.

Empirical results underscore the effectiveness of PersonaAgent, demonstrating its superior performance over baseline methods across four personalization tasks—citation identification, movie tagging, news categorization, and product rating. Notably, PersonaAgent not only achieves high accuracy and F1 scores in classification tasks but also demonstrates low mean absolute error and root mean squared error in regression tasks, indicating its robustness and precision in personalized decision-making scenarios.

Ablation studies validate the vital contributions of each system component; notably, the test-time alignment and persona prompt are critical to achieving precise user modeling. The research affords practical implications by offering a prototype for personalized agents suitable for real-world applications, ranging from tailored educational content to personalized professional assistance. However, the reliance on textual feedback for preference alignment may overlook non-textual user signals, suggesting a potential avenue for future research in multimodal user interactions.

In conclusion, the PersonaAgent framework sets a precedent for embedding personalization within LLM agents. It paves the way for future developments in AI personalization by addressing the challenges associated with dynamic user preferences and real-time adaptability, providing a scalable model that can enhance user-agent interactions in diverse domains.