Skip to main content Skip to navigation

InstructPipe: Generating Visual Blocks Pipelines with Human Instructions and LLMs

Project Overview

The document discusses the application of generative AI in education through the lens of InstructPipe, a visual programming AI assistant designed to simplify the creation of machine learning (ML) pipelines using natural language instructions. By leveraging large language models (LLMs) and a code interpreter, InstructPipe enables users, particularly novices, to generate and visualize pseudocode in a node-graph editor, significantly reducing the number of interactions needed to complete a pipeline and making ML more accessible. The technical evaluation of InstructPipe reveals challenges encountered during user workshops, such as incomplete pipelines and low-quality captions, underscoring the necessity for a rigorous evaluation process. User evaluation methodologies, including semi-structured interviews and counterbalancing, are employed to gain insights into user interactions with the system. Overall, the findings indicate that InstructPipe enhances the onboarding experience in visual programming environments, making it easier for users to engage with complex ML concepts and facilitating a smoother prototyping process.

Key Applications

Visual Programming Tools

Context: Visual programming interfaces designed to assist novice programmers and non-experts in creating digital applications by connecting visual nodes and generating machine learning pipelines through natural language instructions.

Implementation: Implemented as extensions to existing visual programming environments, these tools leverage natural language processing to allow users to create applications and machine learning pipelines visually. Participants interact with the system to design and connect nodes, facilitating the development of programming pipelines without traditional coding.

Outcomes: Significantly reduces user interactions, enhances onboarding experiences, encourages creativity in project prototyping, and facilitates the creation of programming pipelines while reducing the complexity associated with programming for non-expert users.

Challenges: Users may face difficulties with formulating effective prompts, understanding node functionalities, debugging generated pipelines, and encountering incomplete or low-quality outputs during the creation process.

Implementation Barriers

Cognitive Load

Users, particularly novices, find it challenging to articulate their desired pipelines in clear prompts, leading to mental overload. The shift from visual programming to text-based instructions can create confusion and increase cognitive demands.

Proposed Solutions: Future work could focus on developing prompt assistance tools and improved user interfaces that guide users in crafting effective prompts. A multimodal interface that provides visual feedback while users formulate prompts could help bridge the cognitive gap.

Technical Barrier

Participants faced issues such as incomplete pipelines and low-quality captions.

Proposed Solutions: Post-processing was implemented to remove incomplete pipelines and enhance the quality of captions.

Usability Barrier

Non-expert users found some designs and functionalities unintuitive.

Proposed Solutions: Providing in-person assistance during user evaluation to help participants understand the interface.

Project Team

Zhongyi Zhou

Researcher

Jing Jin

Researcher

Vrushank Phadnis

Researcher

Xiuxiu Yuan

Researcher

Jun Jiang

Researcher

Xun Qian

Researcher

Kristen Wright

Researcher

Mark Sherwood

Researcher

Jason Mayes

Researcher

Jingtao Zhou

Researcher

Yiyi Huang

Researcher

Zheng Xu

Researcher

Yinda Zhang

Researcher

Johnny Lee

Researcher

Alex Olwal

Researcher

David Kim

Researcher

Ram Iyengar

Researcher

Na Li

Researcher

Ruofei Du

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Zhongyi Zhou, Jing Jin, Vrushank Phadnis, Xiuxiu Yuan, Jun Jiang, Xun Qian, Kristen Wright, Mark Sherwood, Jason Mayes, Jingtao Zhou, Yiyi Huang, Zheng Xu, Yinda Zhang, Johnny Lee, Alex Olwal, David Kim, Ram Iyengar, Na Li, Ruofei Du

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies