The research paper presents ConstitutionMaker, a tool designed to refine the outputs of large language models (LLM) for creating and customizing chatbots. The tool allows users to provide feedback in natural language, select auto-generated feedback, or rewrite the chatbot’s response. Each mode of feedback generates a principle that guides the chatbot’s behavior. A study with 14 participants showed that ConstitutionMaker made it easier for users to convert their feedback into principles and write them more efficiently. This tool is a step towards interactive critiquing of LLM outputs, helping users to improve chatbot performance.
Publication date: 24 Oct 2023
Project Page: https://arxiv.org/abs/2310.15428v1
Paper: https://arxiv.org/pdf/2310.15428
