Report · Page 70

70
Center for Research on Foundation Models (CRFM)
3.3.2
Foundation models of student thought.
When building AI tools for inclusive, and joyful education, there are many tasks where foundation
models could be useful. Many of those tasks require us to first understand the learners whom we
are trying to help, especially in the context of open ended work.
What would it take for a foundation model to be able to reason about student understanding? It
is easy to imagine a foundation model which has been adapted to answer a math question correctly,
but it is less clear how to build a model that can diagnose mistakes in student understanding based
on the student’s answers. To explore this theme, we consider the case study of providing feedback
to students who are working on open-ended tasks such as writing a short paragraph, drawing a
physics diagram, or writing code. This “feedback challenge” exemplifies how foundation models
can be helpful off-the-shelf for learners, and also demonstrates open areas for foundation model
research.
To effectively provide feedback to students, two central capabilities are required: (1) understand-
ing the subject matter of the task (e.g., physics or coding), and (2) the diagnostic ability to “notice":
a technical term in education for inferring why a student made a mistake. For typical student in-
teractions in a typical classroom, there is not enough data for an AI model to learn, from scratch,
both of these central capabilities. Even for massive courses with millions of students, supervised
algorithms barely understand the complex student reasoning behind even short, four-line programs
[Malik et al. 2021]. As such, the feedback task inherently requires a transfer of understanding from
external data and experience.
Foundation models, as they currently exist, are directly helpful for the first of these capabilities:
understanding a specific subject matter. For example, when learning to provide feedback on short
programming questions, a foundation model such as GPT-3 can efficiently understand what fluent
code looks like with a few examples. Some research in this direction has already started exploring
foundation models that can quickly adapt to questions in new subject matter domains [Wu et al.
2021e; Condor et al. 2021]. Similarly, foundation models could also integrate multiple modes of
information such as the text of a task’s prompt, diagrams in the question, or even the content of a
grading rubric provided to teaching assistants. This unified representational ability can help foun-
dation models comprehend a subject matter through richer sources of information. As a concrete
case study, many of these insights were leveraged as core components of an algorithm which was
able to grade an introductory Computer Science midterm at Stanford University, with the same
effectiveness as human teaching assistants [Wu et al. 2021e]. In this case, subject matter encoding
was built on a foundation model that had been adapted on GitHub code and a corresponding small
dataset for each question’s subject matter. In general, we can imagine leveraging various sources of
data to adapt foundation models to different subject matter. For example, math adaptation could use
mathematical websites or textbooks [Shen et al. 2021b] or historical student answers on platforms
such as Gradescope; spoken language understanding could leverage radio archives or podcasts;
and domains like creative writing could look to large digital archives lik