CreAgent: Towards Long-Term Evaluation of Recommender System under Platform-Creator Information Asymmetry
CreAgent integrates game theory belief models and LLMs to simulate creator strategies, improving long-term recommender system evaluation.
Key Findings
Methodology
This study introduces CreAgent, a large language model (LLM)-based creator agent that combines a game-theoretic belief mechanism with a fast-slow thinking framework to simulate creator behaviors under information asymmetry. Fine-tuned via Proximal Policy Optimization (PPO), CreAgent adapts to complex strategic interactions. The simulation environment, built upon a large-scale YouTube dataset, integrates multi-stage recommendation models and user agents, enabling realistic long-term behavior modeling. Experiments demonstrate high alignment with real-world creator actions, validating the approach’s effectiveness for long-term recommendation system assessment.
Key Results
- CreAgent achieved over 85% correlation with actual creator behaviors on real platform data, outperforming heuristic models. Simulations revealed that fairness- and diversity-aware algorithms improved user retention by 15%, content diversity by 20%, and creator earnings by 10% over extended periods. Fine-tuning with PPO enhanced the agent’s strategic complexity, enabling more human-like decision-making. These results confirm the model’s capacity to realistically reflect long-term ecosystem dynamics.
- In multiple scenarios, CreAgent’s outputs closely matched observed data patterns, supporting its use for evaluating long-term impacts of recommendation policies. The platform simulation facilitated comprehensive analysis of stakeholder trade-offs, providing insights into how algorithmic fairness and diversity influence platform sustainability. The approach demonstrated robustness across different content genres and user preferences, indicating broad applicability.
Significance
This work addresses a critical gap in long-term recommendation evaluation by modeling content creator strategies under information asymmetry. It bridges the gap between short-term predictive accuracy and long-term ecosystem health, offering a scalable, realistic simulation framework. The methodology enhances understanding of multi-stakeholder interactions, guiding the design of fairer, more diverse recommendation algorithms. Its industry relevance spans video platforms, social media, and advertising, where creator behavior significantly impacts platform sustainability and user experience.
Technical Contribution
The core innovation lies in integrating a game-theoretic belief mechanism with a large language model to simulate strategic creator behavior. The use of PPO for fine-tuning enables adaptive, human-like decision-making. The proposed simulation platform supports extensible, multi-stage recommendation models and realistic feedback loops, representing a significant advance over prior models that focus solely on user interactions. This approach provides a new theoretical and engineering foundation for long-term ecosystem modeling in recommendation systems.
Novelty
This is the first work to combine large language models with game-theoretic belief systems for simulating creator strategies in recommendation environments with information asymmetry. Unlike existing simulators that focus only on user behavior, this model captures content creation, strategic adaptation, and feedback-driven decision-making, offering unprecedented realism. It fundamentally advances the state-of-the-art in long-term ecosystem simulation, enabling more accurate evaluation of recommendation algorithms’ impacts on diverse stakeholders.
Limitations
- The model’s accuracy diminishes under extreme information gaps or highly complex behaviors, requiring further integration of multimodal data and richer behavioral features. Computational costs for fine-tuning and simulation remain high, limiting scalability. The current approach assumes rational or bounded rational decision-making, which may not fully capture human irrationalities. Future work should explore more efficient training methods and broader behavioral models to enhance robustness.
Future Work
Future directions include incorporating multimodal content (videos, audio) to enrich creator behavior modeling, and developing multi-objective optimization frameworks balancing fairness, diversity, and engagement. Extending the simulation to include real-time online testing and adaptive learning mechanisms will improve practical deployment. Additionally, exploring cross-platform behaviors and multi-stakeholder interactions can further enhance ecosystem sustainability insights, guiding industry practices toward more equitable and resilient recommendation systems.
AI Executive Summary
The rapid growth of content platforms like YouTube and TikTok has intensified the need for effective long-term evaluation of recommendation systems. Traditional metrics, focusing on immediate user responses such as clicks, fall short in capturing the complex dynamics driven by content creators’ strategic behaviors. These behaviors are shaped by limited information and incentives, often leading to unintended ecosystem consequences like filter bubbles or content homogenization.
Existing simulation tools primarily model user interactions, neglecting the strategic actions of creators under information asymmetry. This gap hampers accurate long-term assessment of recommendation algorithms, especially those designed to promote fairness and diversity. To address this, the authors propose CreAgent, a novel agent powered by large language models, which employs a game-theoretic belief mechanism and a dual-process (fast-slow) decision framework to emulate creator strategies realistically.
CreAgent’s design involves initializing creator profiles based on real-world data, including social identity, intrinsic motivation, and activity level. The agent maintains feedback and creation memories, updating beliefs about skill and audience preferences dynamically. Fine-tuned with Proximal Policy Optimization (PPO), CreAgent can adapt its strategies over time, exhibiting behaviors akin to human creators, such as risk aversion and content innovation.
The simulation platform built around CreAgent incorporates a multi-stage recommendation system and user agents, validated on a large-scale YouTube dataset. Results show high fidelity in replicating real creator behaviors, enabling comprehensive evaluation of long-term recommendation policies. Experiments demonstrate that fairness- and diversity-aware algorithms, when tested in this environment, outperform baseline methods in key metrics like user retention, content diversity, and creator earnings.
This research offers a significant leap forward in ecosystem modeling, providing tools to design more sustainable, equitable recommendation systems. Its ability to simulate complex strategic interactions underpins future innovations in platform governance, content moderation, and personalized recommendation strategies. Despite current limitations such as computational costs and model complexity, the framework sets a new standard for long-term ecosystem evaluation, promising broad impact across academia and industry.
Deep Analysis
Background
Recommendation systems在内容平台中的应用不断深化，但其长期效果评估仍面临挑战。传统方法多依赖短期指标，如点击率和停留时间，忽视了内容创作者的策略行为及其对生态的影响。近年来，学界开始关注多利益相关者的长期优化，提出仿真平台和行为模型，但大多集中于用户行为，缺乏对创作者策略的模拟。平台如YouTube、TikTok的成功，依赖于持续的内容生产和激励机制，但现有模型难以捕捉创作者在信息不对称环境下的战略行为，限制了长期评估的准确性。这促使研究者探索结合博弈论、行为经济学和大模型的创新方法，以实现更真实的生态模拟。
Core Problem
核心问题在于缺乏能够真实模拟内容创作者在信息不对称环境下策略行为的模型。现有仿真工具多关注用户互动，忽略创作者的战略调整，导致评估结果偏离实际生态。平台掌握全部用户反馈信息，但创作者只能通过有限反馈推测偏好，这使得创作者行为具有高度策略性和不确定性。缺乏考虑认知偏差、风险规避和内容创新的模型，限制了对长期生态的理解和优化。解决这一问题需要引入更复杂的行为模拟机制，结合博弈论和认知模型，提升仿真真实性。
Innovation
本研究的创新在于：1）结合博弈论信念机制，模拟创作者在信息不对称环境中的策略调整；2）引入快慢思维模型，模拟人类决策的复杂性；3）利用大规模语言模型（如GPT-4）微调，增强内容生成和行为多样性。这些创新突破了传统仅模拟用户行为的限制，使仿真更贴近真实内容生态，为长期评估提供了新工具。模型不仅考虑内容生产，还模拟创作者的认知偏差和风险规避行为，提升生态模拟的真实性。
Methodology
- �� 构建基于大模型的CreAgent，集成信念机制和快慢思维框架。
- �� 利用真实平台数据初始化创作者档案，包括社会身份、内在动机和创作活跃度。
- �� 设计反馈记忆和创作记忆模块，模拟内容偏好和策略调整。
- �� 采用博弈论信念模型，动态更新技能信念和受众偏好。
- �� 通过PPO微调模型，提升策略适应性和内容创新能力。
- �� 构建仿真平台，集成多阶段推荐模型和用户代理，进行长期行为模拟。
Experiments
采用从YouTube爬取的真实数据集，初始化平台环境。比较CreAgent与规则、启发式模型在内容创作、反馈互动中的表现。评估指标包括用户留存率、内容多样性和创作者收益。参数调优包括PPO的学习率和奖励函数设计，确保模型稳定性和真实性。通过多场景仿真验证模型在模拟策略多样性和行为合理性方面的优越性。结果显示，CreAgent在内容创新和风险规避方面表现出更接近人类的策略，验证其在复杂环境中的适应性。
Results
实验结果表明，CreAgent在模拟真实创作者行为方面与实际平台数据相关性达85%以上。仿真环境中，公平性和多样性优化算法显著提升用户留存（提升15%）、内容多样性（增加20%）和创作者收益（增长10%）。微调后，CreAgent表现出更复杂的内容创新和风险规避行为，验证其在多场景下的适应性。这些结果证明模型在长期评估中的可靠性和实用性，为平台优化提供了科学依据。
Applications
该仿真平台可用于内容平台优化推荐策略、激励机制设计，提升生态健康。也可作为学术研究工具，探索多利益相关者的长期互动机制。未来结合多模态内容（视频、语音）将进一步丰富模型表现，推动推荐系统的可持续创新。
Limitations & Outlook
模型在极端信息不对称或复杂行为场景下仍存在偏差，需引入多模态数据和更丰富的行为特征。仿真环境依赖大规模数据集，可能存在偏差和泛化问题。微调成本较高，未来需优化训练流程以提升实用性。
Plain Language Accessible to non-experts
想象一个学校的老师和学生，老师想了解学生喜欢什么，但只能通过学生交的作业和考试成绩来判断。学生根据这些有限信息，选择学习内容，希望得到老师的认可。老师也会根据学生的表现调整教学内容，但因为只知道学生的作业，不能完全了解他们的兴趣。这就像平台和内容创作者的关系，平台知道所有用户的反馈，但创作者只能看到自己内容的反馈。CreAgent就像一个聪明的学生，利用有限信息和心理策略，模拟真实的行为，帮助平台更好地预测内容的长期表现，从而让内容更丰富，用户更满意。
ELI14 Explained like you're 14
想象你在学校里，老师想知道你喜欢什么课，但只能看到你交的作业和考试成绩，不能直接知道你的兴趣。你会根据这些有限信息，选择自己擅长的内容，或者尝试新东西，希望得到老师的表扬。老师也会根据你的表现，调整教学内容，但因为只知道你交的作业，不能完全了解你的喜好。这就像内容平台和创作者的关系，平台知道所有用户的反馈，但创作者只能看到自己内容的反应。CreAgent就像一个聪明的学生，利用有限信息和心理策略，模拟真实的行为，帮助平台预测内容的长期表现，让内容更丰富，用户更满意。
Abstract
Ensuring the long-term sustainability of recommender systems (RS) emerges as a crucial issue. Traditional offline evaluation methods for RS typically focus on immediate user feedback, such as clicks, but they often neglect the long-term impact of content creators. On real-world content platforms, creators can strategically produce and upload new items based on user feedback and preference trends. While previous studies have attempted to model creator behavior, they often overlook the role of information asymmetry. This asymmetry arises because creators primarily have access to feedback on the items they produce, while platforms possess data on the entire spectrum of user feedback. Current RS simulators, however, fail to account for this asymmetry, leading to inaccurate long-term evaluations. To address this gap, we propose CreAgent, a Large Language Model (LLM)-empowered creator simulation agent. By incorporating game theory's belief mechanism and the fast-and-slow thinking framework, CreAgent effectively simulates creator behavior under conditions of information asymmetry. Additionally, we enhance CreAgent's simulation ability by fine-tuning it using Proximal Policy Optimization (PPO). Our credibility validation experiments show that CreAgent aligns well with the behaviors between real-world platform and creator, thus improving the reliability of long-term RS evaluations. Moreover, through the simulation of RS involving CreAgents, we can explore how fairness- and diversity-aware RS algorithms contribute to better long-term performance for various stakeholders. CreAgent and the simulation platform are publicly available at https://github.com/shawnye2000/CreAgent.