Hi Neil,

Thanks for asking. We hypothesize that there would be a sweet spot for the prior knowledge (the readiness of teaching) that would most beneficial for learning and teaching, and understanding where would be an important research question. We do not have an answer just yet. A previous study where we compared learning by teaching and cognitive tutoring (aka learning by being tutored) showed that students who have low prior knowledge would learn more with cognitive tutoring. To answer where the sweet spot would be, we plan to extend our current software so that students can freely switch between learning by teaching and cognitive tutoring. That way, we would be able to measure the impact of prior knowledge on tutor learning.

SimStudent is a genuine machine-learning agent that does learn skills when being tutored. SimStudent is a cognitive model of inductive learning, and a previous study showed that SimStudent has a cognitive fidelity in a sense that it makes same errors that students commonly make; e.g., subtracting 4 from both sides of 2x-4 = 10. SimStudent make such an error due to an induction error when “shallow” prior knowledge is given—e.g., perceiving ‘4’ in 2x-4 as a “number” instead of perceiving ‘-4’ as a term where such perception would work to, say, subtract 5 from 3x+5 = 9.

We haven’t compared learning by teaching synthetic peer vs. real peer. An interesting future study!

Hi Jessica,

Thanks for great questions!

SimStudent attempts to solve a problem given by the student (who acts as a “tutor” for SimStudent). SimStudent asks “why” questions in the context of tutoring. It asks “why am I wrong?” when the student disagrees a step suggested by SimStudent. It asks the reason why the student suggested a particular step (e.g., “why did you divide both sides by 4”) when SimStudent got stuck and asked a help to the student; and the student suggested a next step.

The theory says that when the tutee asks a constructive question, then the effect of tutor learning would increase. We are currently improving SimStudent’s ability to ask questions by integrating a natural language processing component to SimStudent so that SimStudent recognizes “good” and “bad” explanations.

A previous study showed that when Mr. Williams provided metacognitive scaffolding on what problem should be taught then students tend to provide more appropriate problems to SimStudent, which results in increasing the change of passing quiz (i.e., increasing SimStudent’s proficiency) that further increases student’s learning.

Hi Jermy,

Nice question! In classroom studies, we have observed lots of students indeed believe SimStudent does understand their response hence they are engaged in typing responses. (at the same time, some students realize that SimStudent doesn’t really understand their input, which is true). We hypothesize that the rich dialogue between students and SimStudent would facilitate students’ conceptual understanding, which is a benefit of learning by teaching that does not happen in the context of cognitive tutoring (i.e., learning by being tutored).

Hi Stephanie,
Thanks for asking!
The demo you see is what students do. You, as a student, enter a problem; you quiz SimStudent; you answer SimStudent’s “why” questions. The interaction between you and SimStudent is the key component of learning by teaching.
Yes, in the classroom studies, we occasionally observe students “gaming” the system—e.g., students enter silly response to SimStudent’s questions. We also observe suboptimal behaviors—e.g., students teach SimStudent only on “easy” problems that they can teach, which results in SimStudent’s poor learning (hence student’s poor learning as well).
We have indeed started a collaboration with Ryan Baker and his colleagues to implement suboptimal behavior detector (including “gaming” the system) to provide the proactive scaffolding to students to enhance learning by teaching.