The paper "Large Language Models are Learnable Planners for Long-Term Recommendation" addresses the challenge of enhancing long-term engagement in recommendation systems by leveraging the planning capabilities of Large Language Models (LLMs). Traditional methods often focus on immediate user responses, leading to issues such as filter bubbles and a lack of long-term engagement. To tackle this, the authors propose a Bi-level Learnable LLM Planner (BiLLP) framework, which integrates macro-learning and micro-learning mechanisms to improve LLMs' planning abilities. - **Key Contributions:** - Introduces the concept of using LLMs for long-term planning in recommendations. - Proposes a BiLLP framework that enhances LLMs' planning capabilities through macro-learning and micro-learning. - Conducts extensive experiments to validate the effectiveness of the BiLLP framework. - **Methodology:** - **Macro-Learning:** Involves a Planner and Reflector module that generates high-level plans and reflects on past experiences to update the Planner's memory. - **Micro-Learning:** Involves an Actor-Critic module that personalizes high-level plans into executable actions and evaluates their effectiveness. - **Experiments:** - Evaluates the BiLLP framework using simulated environments and compares it with state-of-the-art reinforcement learning (RL) methods and other LLM-based approaches. - Demonstrates that BiLLP outperforms existing methods in terms of trajectory length, average single-round reward, and cumulative reward, indicating superior long-term engagement. - **Results:** - BiLLP consistently achieves better performance in both trajectory length and cumulative reward, showing its effectiveness in mitigating filter bubbles and maximizing long-term engagement. - Ablation studies confirm the importance of both macro-learning and micro-learning mechanisms. - The Critic module is shown to effectively estimate the state-value function, improving the Actor's policy updates. - The framework is robust across different recommendation environments and base LLM models. - **Conclusion:** - The paper concludes by highlighting the potential of integrating planning capabilities from LLMs into recommendation systems to optimize long-term engagement. - Future research directions include enhancing small-scale models and integrating reinforcement learning algorithms within the planning framework.The paper "Large Language Models are Learnable Planners for Long-Term Recommendation" addresses the challenge of enhancing long-term engagement in recommendation systems by leveraging the planning capabilities of Large Language Models (LLMs). Traditional methods often focus on immediate user responses, leading to issues such as filter bubbles and a lack of long-term engagement. To tackle this, the authors propose a Bi-level Learnable LLM Planner (BiLLP) framework, which integrates macro-learning and micro-learning mechanisms to improve LLMs' planning abilities. - **Key Contributions:** - Introduces the concept of using LLMs for long-term planning in recommendations. - Proposes a BiLLP framework that enhances LLMs' planning capabilities through macro-learning and micro-learning. - Conducts extensive experiments to validate the effectiveness of the BiLLP framework. - **Methodology:** - **Macro-Learning:** Involves a Planner and Reflector module that generates high-level plans and reflects on past experiences to update the Planner's memory. - **Micro-Learning:** Involves an Actor-Critic module that personalizes high-level plans into executable actions and evaluates their effectiveness. - **Experiments:** - Evaluates the BiLLP framework using simulated environments and compares it with state-of-the-art reinforcement learning (RL) methods and other LLM-based approaches. - Demonstrates that BiLLP outperforms existing methods in terms of trajectory length, average single-round reward, and cumulative reward, indicating superior long-term engagement. - **Results:** - BiLLP consistently achieves better performance in both trajectory length and cumulative reward, showing its effectiveness in mitigating filter bubbles and maximizing long-term engagement. - Ablation studies confirm the importance of both macro-learning and micro-learning mechanisms. - The Critic module is shown to effectively estimate the state-value function, improving the Actor's policy updates. - The framework is robust across different recommendation environments and base LLM models. - **Conclusion:** - The paper concludes by highlighting the potential of integrating planning capabilities from LLMs into recommendation systems to optimize long-term engagement. - Future research directions include enhancing small-scale models and integrating reinforcement learning algorithms within the planning framework.