Knowledge Boundary and Persona Dynamic Shape A Better Social Media Agent

TL;DR

The paper tackles two core problems in social media agents: knowledge boundary, ensuring the agent's world knowledge stays aligned with its persona, and persona dynamics, ensuring only action-relevant persona information is used. It proposes a two-part framework comprising agent simulation and system construction, integrating personalized knowledge from external sources (e.g., HotpotQA) with dynamic internal persona retrieval, implemented via a five-module agent (persona, action, planning, memory, reflection) and a Mastodon-based sandbox with a recommendation mechanism. Key contributions include the design of knowledge-bound personalization, dynamic retrieval to minimize interference, a text-based social media sandbox, and comprehensive automatic and human evaluations showing improved action rationality and text quality. The work advances personalized, anthropomorphic social agents and provides a practical evaluation environment, with future work aiming at multimodal inputs and expanded knowledge sources, guided by thresholds $T_k$ and $T_p$ in the personalization workflow.

Abstract

Constructing personalized and anthropomorphic agents holds significant importance in the simulation of social networks. However, there are still two key problems in existing works: the agent possesses world knowledge that does not belong to its personas, and it cannot eliminate the interference of diverse persona information on current actions, which reduces the personalization and anthropomorphism of the agent. To solve the above problems, we construct the social media agent based on personalized knowledge and dynamic persona information. For personalized knowledge, we add external knowledge sources and match them with the persona information of agents, thereby giving the agent personalized world knowledge. For dynamic persona information, we use current action information to internally retrieve the persona information of the agent, thereby reducing the interference of diverse persona information on the current action. To make the agent suitable for social media, we design five basic modules for it: persona, planning, action, memory and reflection. To provide an interaction and verification environment for the agent, we build a social media simulation sandbox. In the experimental verification, automatic and human evaluations demonstrated the effectiveness of the agent we constructed.

Figures (12)

• Figure 1: Knowledge boundary and persona dynamic shape personalization in social media: (a) knowledge boundary: the agent should only have world knowledge that is consistent with the persona; (b) persona dynamic: the agent should only use personal information related to the current action when performing actions.
• Figure 2: The overall framework of our work is divided into two parts: system construction and agent simulation. The system construction part consists of the basic functions of social media, the recommendation mechanism and the simulation of overall user activity. The agent simulation part consists of five modules: persona, planning, action, memory and reflection. In the persona module, the agent can obtain personalized knowledge from external knowledge sources, and the persona information is internally retrievable when performing actions.
• Figure 3: An example of the process of enriching persona information. The text marked in blue is the basic information of the persona. The text marked in red is the advanced attributes of the persona.
• Figure 4: The prompt for the agent to generate post content. Text marked in red is personalized knowledge retrieved using post topics. The text marked in blue is the persona information obtained through internal retrieval.
• Figure 5: An example of the generated planning content.