Do Generative Recommenders Deepen the Information Cocoon? A Closed-Loop Simulation with LLM-powered User Simulators

In an era overwhelmed by information, recommendation systems serve as vital gatekeepers, guiding users through vast content landscapes. However, this guiding role can inadvertently lead to the formation of 'information cocoons'—self-reinforcing loops where users are repeatedly exposed to similar content, narrowing their horizons and reinforcing existing preferences. Traditional recommendation models, especially those based on atomic item IDs, have long been scrutinized for their propensity to foster such content homogeneity. Yet, with the advent of generative recommendation systems that encode items as discrete code sequences (SID sequences) and generate recommendations autoregressively, a new question arises: do these models deepen or mitigate the cocoon phenomenon?

Addressing this, the authors introduce RecLoop, a sophisticated closed-loop simulation framework that leverages large language models (LLMs) to emulate user behavior over extended feedback cycles. This framework enables a detailed examination of how generative recommenders influence content diversity over time. By simulating 15 feedback cycles on two Amazon datasets—Office Products and Toys & Games—with thousands of virtual users, the study offers a comprehensive analysis of long-term cocoon dynamics. The user simulators maintain dynamic preferences, updating their states as they interact with recommendations, thus capturing realistic preference evolution.

A key innovation is the introduction of the Code-Space Structural Cocoon metric, which quantifies how concentrated the generated code representations become across feedback cycles. This allows the authors to assess whether the internal generative space itself narrows over time, complementing traditional exposure-level metrics. The experimental results reveal that generative recommenders generally outperform traditional ID-based models in maintaining exposure diversity, thereby slowing the homogenization process across users. Larger models, with more parameters, further enhance diversity preservation, indicating that model capacity plays a crucial role.

However, the study also uncovers that feedback loops can induce concentration within the generated code space, especially under certain tokenization strategies. Collaborative-signal tokenization tends to produce stronger cocoon effects than semantic tokenization, highlighting the importance of encoding choices. These findings suggest that the design of tokenization strategies and the scale of models are critical factors influencing content diversity in generative recommendation systems.

Overall, this research advances our understanding of the long-term behaviors of generative recommenders, demonstrating their potential to better preserve content diversity compared to traditional models. It also emphasizes the importance of internal model representations and encoding strategies in shaping cocoon effects. The insights gained pave the way for future innovations aimed at designing recommendation systems that balance personalization with content variety, ultimately fostering healthier, more inclusive information ecosystems. Despite some limitations—such as reliance on simulated users and dataset scope—this work provides a foundational step toward more transparent and controllable generative recommendation architectures.