Core Concepts
DirectGPT enhances interaction with large language models through direct manipulation principles.
Abstract
The content discusses the implementation of DirectGPT, a user interface layer on top of ChatGPT that transforms direct manipulation actions into engineered prompts. It focuses on improving interactions with large language models by providing continuous representation of objects, reusing prompt syntax, manipulable outputs, and undo mechanisms. A study showed improved efficiency and effectiveness compared to baseline ChatGPT.
Abstract:
Principles of direct manipulation improve interaction with large language models.
Continuous representation of generated objects.
Reuse of prompt syntax in toolbar commands.
Manipulable outputs to control prompts' effects.
Undo mechanisms for reversible operations.
Introduction:
Direct manipulation interfaces emerged as an alternative to command line interfaces.
Current interfaces for LLMs lack benefits like improved learnability and speed due to indirect engagement.
Background and Related Work:
Direct manipulation principles defined by Shneiderman.
Issues with prompting in LLMs motivate the use of direct manipulation.
Systems proposed to help craft better prompts for LLMs.
DirectGPT: An Exemplar Direct Interface for LLMs:
Describes how DirectGPT implements direct manipulation principles.
Illustrates the utility of DirectGPT through a use case scenario.
Stats
データ、コード、およびデモはhttps://osf.io/3wt6sで利用可能です。