TAIRA: Thought-Augmented Planning for LLM-Powered Interactive Recommender Agent

Thought-Augmented Planning for LLM-Powered Interactive Recommender Agent Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '26)

Haocheng Yu, Yaxiong Wu, Hao Wang, Wei Guo, Yong Liu, Yawen Li, Yuyang Ye, Junping Du, and Enhong Chen.

---

📖 Introduction

Interactive recommender systems (IRS) allow users to express needs via natural language. However, existing LLM-powered agents often struggle with complex, unrefined, or ambiguous user intents due to limited planning and generalization capabilities.

We propose TAIRA (Thought-Augmented Interactive Recommender Agent), a novel multi-agent system designed to handle complex user intents through:

1. Thought Pattern Distillation (TPD): A mechanism that distills high-level reasoning patterns from both successful agent trajectories and expert-corrected failures.
2. Hierarchical Planning: A manager agent orchestrates tasks by decomposing user needs and dynamically updating plans based on execution feedback.
3. Generalization: Leveraging distilled thought patterns to solve novel tasks effectively.

Figure 1: Examples of recommendation involving diverse and complex user intent and thought-augmentation with past experiences.

🚀 Framework Architecture

TAIRA operates as a multi-agent system featuring a Manager Agent that orchestrates specialized Executor Agents (Searcher, Item Retriever, etc.). The system continuously learns and refines its planning capabilities through the Thought Pattern Distillation module.

Figure 2: The overall architecture of TAIRA, illustrating the Manager Agent's planning process, Executor Agents, and the TPD mechanism.

🛠️ Installation

Prerequisites

• Python 3.12.7 or higher
• PyTorch (Check official site for your CUDA version)

Setup

1. Clone the repository:

   git clone [https://github.com/Alcein/TAIRA.git](https://github.com/Alcein/TAIRA.git)
   cd TAIRA

2. Install dependencies:

   pip install -r requirements.txt

⚙️ Configuration (system_config.yaml)

Before running the program, you need to configure the system_config.yaml file based on your specific requirements.

Explanation of the Parameters:

• QUERY_NUMBER: Limits the size of the dataset. Here, it's set to 500, meaning it will process up to 200 data points.
• TOPN_ITEMS: Number of top items returned, set to 500.
• TOPK_ITEMS: The number of top K items selected, set to 10.
• DOMAIN: The data domain being used. Options include "amazon_clothing", "amazon_beauty", and "amazon_music". The default is set to "amazon_clothing".
• MODEL: Specifies the model to be used. The default model is "gpt-4o", but you can switch to other models.
• METHOD: The method being used. In this case, it's set to TAIRA.
• OPENAI_BASE_URL and OPENAI_API_KEY: These fields are for configuring access to the OpenAI API. Make sure to provide your valid OpenAI API key. If you need to use a model other than openai, please set the corresponding base_url and api_key.
• GOOGLE_API_KEY and GOOGLE_CSE_ID: These are for Google API configurations in Searcher Agent.

Ensure you have a valid OpenAI API key set up in the OPENAI_API_KEY field for the program to work properly.

🏃 Usage

To start the interactive recommendation simulation and evaluation:

python main.py