PersonaAct: Simulating Short-Video Users with Personalized Agents for Counterfactual Filter Bubble Auditing

TL;DR

PersonaAct tackles filter bubbles in short-video recommendations by learning persona-conditioned multimodal agents from real traces to enable scalable, privacy-preserving auditing. It combines automated persona interviews with a two-stage training regime (supervised fine-tuning followed by reinforcement learning) to produce realistic behavior. The framework enables counterfactual auditing across platforms by measuring content diversity (bubble breadth) and escape potential (bubble depth) through BEP, revealing platform-specific inertia and significant content narrowing. The authors release the first open multimodal short-video dataset and code, facilitating reproducible auditing of recommender systems and providing a foundation for future mitigation strategies.

Abstract

Short-video platforms rely on personalized recommendation, raising concerns about filter bubbles that narrow content exposure. Auditing such phenomena at scale is challenging because real user studies are costly and privacy-sensitive, and existing simulators fail to reproduce realistic behaviors due to their reliance on textual signals and weak personalization. We propose PersonaAct, a framework for simulating short-video users with persona-conditioned multimodal agents trained on real behavioral traces for auditing filter bubbles in breadth and depth. PersonaAct synthesizes interpretable personas through automated interviews combining behavioral analysis with structured questioning, then trains agents on multimodal observations using supervised fine-tuning and reinforcement learning. We deploy trained agents for filter bubble auditing and evaluate bubble breadth via content diversity and bubble depth via escape potential. The evaluation demonstrates substantial improvements in fidelity over generic LLM baselines, enabling realistic behavior reproduction. Results reveal significant content narrowing over interaction. However, we find that Bilibili demonstrates the strongest escape potential. We release the first open multimodal short-video dataset and code to support reproducible auditing of recommender systems.

Figures (5)

• Figure 1: RecSys auditing approaches: cost, privacy, fidelity trade-offs.
• Figure 2: Data collection interface mirroring mobile device screen.
• Figure 3: Overview of PersonaAct. We first collect multimodal browsing traces from real users. An automated interview then integrates behavioral analysis with structured questioning to produce an interpretable persona. Next, we train a persona-conditioned agent to simulate user actions from multimodal context. Finally, the trained agents are deployed for filter bubble auditing by tracking content diversity and escape potential over time.
• Figure 4: Automated Persona Interview example.
• Figure 5: Content diversity over interaction.