Teachable agent

What is it?

In the teachable agent approach, a human teaches a topic to an artificial student. AlgoBo, in TeachYou, uses a flow that regulates the knowledge state and response behavior; it is more comprehensive than just the sentence “pretend you don't know”. The model's weights are not trained by this chat.

When does it help?

For noticing the knowledge you assume and the gaps in your explanation while teaching a concept.

Examples

The situations and responses below are fictional teaching examples; they are not results from a model, tool or benchmark that was actually run.

Simple

Situation

You will explain to an artificial student what a loop does.

Prompt

Human teacher; model student simulation. This is not the full TeachYou system but a limited chat adaptation.
Knowledge state: You know lists; you don't know loops yet. Do not reason as if you knew concepts I have not explained to you. My explanation: "A loop does the same thing for each item in the list."
Say in one sentence only what you understood from this explanation; then ask a single why/how question. Wait for my answer. In at most two rounds, let's summarize together what changed in the knowledge state.

Sample output

Student: “The same operation is applied to each item. How many times does it run if the list is empty?” Human: “It doesn't run at all.”

What did we get?

The person teaching had to add the boundary case to their explanation.

Medium

Situation

The student simulation must not act as if it knows more than the teacher said.

Prompt

Application state: known=[list, loop], open_concepts=[condition], taught=[].
Human explanation: "To pick the positive numbers, we check whether each number is greater than zero."
Stage 1, a separate model call: Make a short state proposal of which concept this explanation taught and which uncertainty it left.
The supervisor should store the approved state update. Stage 2, a separate response call: Give a student answer using only this state; check with a question whether zero is included.
After the human's answer, one more round can be done; stop at 4 model calls in total. Do not claim to remember state that was not saved.

Sample output

Student: “I'm taking the ones greater than zero; is zero left out?” The teacher's answer is added to the next knowledge record.

What did we get?

The student role was bounded by a visible state; where the knowledge came from was tracked.

Hard

Situation

The teacher has a wrong generalization; a helper channel should support the teaching.

Prompt

Design a stateful teachable agent and a separate teaching helper. Topic: finding the largest number in a list. Human: "The largest value starts at 0."
The student response call should briefly say how it would handle the list [-5,-2] with the rule it has learned; it should not give the fully correct solution from outside knowledge.
A separate helper call should examine the conversation and the topic rule; it should give the teacher only a suggested counterexample question. Do not present the helper message as the student's answer.
When the human corrects the explanation, the state should be updated; a teaching check should be done with a new negative list. At most two teaching rounds; do not declare a learning gain without a final independent human exercise.

Sample output

If the student heads toward the result 0, the helper suggests the question to the teacher: “If all the numbers are negative, what should the starting value be?”

What did we get?

The error in the teacher's explanation and the helper's role were seen in separate channels.

Where should you stop?

A model talking like a student is not a real independent student or weight training. AlgoBo's finding about knowledge-building conversation should not be confused with a post-test gain. The learning results of a separate music education study are not the results of this coding system either.

Sources

• Teach AI How to Code: Using Large Language Models as Teachable Agents for Programming Education (new tab) — Jin, Hyoungwook; Lee, Seonghee; Shin, Hyungyu; Kim, Juho. 2023-09-25; version read 2024-03-11. Defines the TeachYou/AlgoBo flow that limits the knowledge state and regulates Reflect–Respond and questioning behavior; conversation measures are not the same as a learning test. Evidence level: relevant body sections of the original paper.

  Copy source address

• Exploring the Impact of an LLM-Powered Teachable Agent on Learning Gains and Cognitive Load in Music Education (new tab) — Jin, Lingxi; Lin, Baicheng; Hong, Mengze; Zhang, Kun; So, Hyo-Jeong. 2025-04-01. A teachable agent study in a separate music education context; its results do not carry over to AlgoBo's coding study. Evidence level: relevant body sections of the original paper.

  Copy source address