Using agentic workflows to build AI-integrated assignments for introductory psychology

The Applied Social Cognition (ASC) Lab from the School of Psychological Science would like to develop AI-integrated assignments for the introductory psychology classroom using agentic workflows. Each step of the workflow will be handled by a different LLM (or possibly the same LLM but re-prompted independently at each stage) or other automated process. The output of one step will be fed as input into the next, with separate prompts designed for each specific stage.

A simple example: Instead of prompting a single LLM with a multi-step process, "First, generate ten topic ideas. Next, choose the best one. Then, create an outline based on that topic. Finally, use the outline to write a three-paragraph essay." With an agentic workflow, the task would be broken up into separate steps, with each step handled by a separate LLM instance. For this example, one LLM might generate the topics, another might choose the best one, a third might create the outline, and a fourth could write the essay. In this approach, each step would have its own unique prompt, and the output of each step would be fed into the input of the next step. 

A slightly more relevant example: From a student's point of view, imagine that an AI-integrated assignment first presents you with learning content, and then asks you to summarize that content in your own words. An assignment workflow LLM could then analyze your answer and determine whether or not it displayed sufficient understanding of the learning content. If your response does display sufficient understanding, then the workflow will present new learning content. If your response does not display sufficient understanding, then the workflow will re-explain the original learning content and then ask you another question to assess your understanding. This iterative process could proceed through an entire textbook chapter section of content. 

Capstone students will first familiarize themselves with the Retrieval-Augmented Generation system implemented by a 2023-2024 Capstone team. We will then design workflows which use multiple LLMs to build interactive classroom assignments which guide students through multi-step processes of learning. We will be working with GPT-4o and Llama 3 APIs, Retrieval-Augmented Generation (RAG) techniques, a Qdrant vector database, and our final products will be hosted on Amazon EC2. 

We will deploy these assignments within actual intro psych classes in Winter 2025. They will be accompanied by user experience surveys and we will collect the chat logs of the student-LLM interactions. In Spring 2025, we will analyze the user experience and chat log results, implement changes in our workflows, and submit our findings to a conference. 

We are looking for 3-4 highly motivated students.

Objectives


  • Design and deploy at least one AI-integrated assignment into the Winter 2025 intro psych classroom
  • Make use of user experience questionnaires and chat logs to modify workflow
  • Write up results and submit to a conference

Motivations


Traditional classroom settings often struggle to cater to the individual learning needs of each student. LLM-powered assignments can offer personalized instruction for every student, adapting its teaching style and content to each student's learning curve. Personalized tutoring can lead to increased student engagement, which is a strong predictor of academic success. Working with AI assignments can also improve students' digital literacy. If successful, our workflow could serve as a template for other subjects and education levels, offering a scalable solution to personalized education. This project would place OSU at the forefront of educational technology research. 

Qualifications


Minimum Qualifications:
None Listed

Preferred Qualifications:

Requisite coding knowledge (mostly Python) to learn how to work with LLM APIs and RAG procedures


Details


Project Partner:

Joseph Slade

NDA/IPA:

No Agreement Required

Number Groups:

1

Project Status:

Accepting Applicants

Keywords:
Operations / CloudResearchMachine Learning (ML)Education
Card Image Capstone