RecPrompt is the first prompt engineering framework that integrates a series of intermediate reasoning steps for news recommendation, leveraging the capabilities of LLMs to perform complex news recommendation tasks. This framework incorporates a prompt optimizer that applies an iterative bootstrapping process, significantly improving the alignment of news content with user preferences.

To set up the project, you should first create a JSON file under config/ named api_keys.json with the following API_KEY:

{
  "OPENAI_API_KEY": "YOUR OPEN API KEY",
  "HF_TOKEN": "YOUR HUGGINGFACE TOKEN"
}

Running the framework requires the following steps:

1. Install the required libraries by running pip install -r requirements.txt
2. Execute the RecPrompt optimizing script by running python framework/run_rec_prompt.py --tuning_list=3tune3_io,4tune3_io
3. Execute the recommendation task to reproduce the results by running python framework/run_rec_test.py
4. Execute evaluation of TopicScore by running python evaluation/evaluate_news_topics.py

The following table shows the tuning procedure for the RecPrompt framework using GPT-3.5 as the news recommender. The framework is tuned using the IO and CoT prompts with GPT-3.5 and GPT-4 as prompt optimizer.

Template	Tuned by GPT-3.5	Tuned by GPT-4
IO
CoT