PersonaAgent: When Large Language Model Agents Meet Personalization at Test Time
Weizhi Zhang, Xinyang Zhang, Chenwei Zhang, Liangwei Yang, Jingbo Shang, Zhepei Wei, Henry Peng Zou, Zijie Huang, Zhengyang Wang, Yifan Gao, Xiaoman Pan, Lian Xiong, Jingguo Liu, Philip S. Yu, Xian Li
TL;DR
PersonaAgent is a novel framework that personalizes large language model agents at test time by integrating memory modules and user-specific personas, enabling real-time adaptation to user preferences and improving task performance.
Contribution
It introduces the first personalized LLM agent framework with test-time user preference alignment, combining memory modules and persona prompts for dynamic personalization.
Findings
Outperforms baseline methods in personalization accuracy
Effectively scales in real-world test-time scenarios
Enhances user experience through tailored responses
Abstract
Large Language Model (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…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. PersonaAgent is the first LLM agent framework for dynamic user-level personalization, combining episodic/semantic memory and persona-driven actions for continuous adaptation. 2. The test-time alignment method optimizes the persona via simulated interactions and textual loss, enabling real-time, scalable personalization without retraining. 3. The work rigorously validates its approach across four diverse personalization tasks, ablation studies, and scaling analyses.
1. The evaluation relies primarily on LaMP, which focuses on text classification and generation tasks that do not adequately capture instruction-following ability in real interactive dialogues—would the framework still excel other benchmarks? 2. The action module only uses Wikipedia search and personal data retrieval; given that Wikipedia search may dominate performance gains, does the personalization component (i.e., personal data retrieval alone) meaningfully contribute to the agent’s effectiv
* Proposes a unified memory-action framework for personalization, generalizable across tasks. * Introduces a test-time persona optimization mechanism, enabling real-time adaptation to user preferences. * Provides comprehensive experiments and ablation studies, showing the necessity of each component.
* Evaluation relies on machine metrics (accuracy, F1, ROUGE) not fully convincing; would be better to include personalization metrics (e.g., Persona-F1, faithfulness). * The computational cost and scalability of test-time alignment are not thoroughly discussed.
1. Using the persona as an intermediary between memory and action is intuitive and reasonable. 2. A complete personalized agent framework, PersonaAgent, is proposed. 3. Experimental results demonstrate the effectiveness of the proposed strategy, and detailed ablation experiments are performed.
1. The paper is poorly written, the method is obscure, and lacks necessary details. For example, the input to $f_{enc}$ is a tuple. How is the tuple encoded into an embedding? How does the resulting $\mathcal{R}^u(q^*)$ work? What is the observation? How are personas and observations combined to perform personalization? What is the textual loss function? These are unclear. 2. The motivation for the proposed module is unclear, making this paper less like a technical report. 3. Are the baseline me
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsPersona Design and Applications · Topic Modeling · Multimodal Machine Learning Applications
MethodsADaptive gradient method with the OPTimal convergence rate