Why pedagogy is hard for a language model

A strong language model does what it was trained to do: resolve the question as helpfully as possible. For learning a language, that is often the wrong move.

Answering is not teaching

Good teachers hold back and let a small amount of difficulty do its work. A model's defaults pull the other way.

• "Is it dem or den?" Gets the answer, and has learned only the answer.
• "Which case does mit take?" Finds it, and has learned something that transfers.

The assistance dilemma

The model cannot see where the learner is, unless the system around it shows it.

What a model lacks

• A learner model. What this person recognises, can retrieve and can transfer, per word and per rule.
• An objective. Which CEFR can-do statement the exercise serves, and which depth it tests.
• A sequence. What came before, the prerequisites, and the next step.

What goes wrong without them

• Over-explaining. A paragraph where one line would do.
• Words too hard. Vocabulary two levels above the learner.
• Five corrections at once. Every mistake in one go.
• Shifting names. The same thing called three names across sessions.
• Praise for everything. Which quickly means praise for nothing.

A capable model is one ingredient. Good teaching depends on the decisions around it.

Pedagogy around the model

Feedback in four parts

1. What was wrong One mistake, named.
2. The rule The reason behind the form.
3. One example The rule in a sentence.
4. One thing to try next A single step, not a list.

Tested before learners see it

1. Plant a weakness Synthetic learners with known gaps.
2. Run the change Before it reaches a real learner.
3. Check the finding Does the system find what we planted?