Try this first
Six places in a small town
Work it out in your head, then tap your answer. No drawing yet.
Now draw it.
Tap a square to place the highlighted landmark.
A sketch a language model can draw.
A language model writes only text. To study how it sketches, we put the map in text form: an ASCII grid with one letter per place.
People sketch to think about space. Do language models?
Once a model has learned to sketch, does it reason better from words alone?
It practices one thing: drawing.
“A is above B. C is at 3 o'clock from B. D is northwest of A.”
+--+--+--+ |D | | | +--+--+--+ | |A | | +--+--+--+ | |B |C | +--+--+--+
Learning to draw ASCII from textual spatial descriptions.
Then it gets a question in words.
“D is below E. F is to the left of E. Where is D relative to F?”
- No ASCII given
- No ASCII drawn
It reads the text and answers with a word.
And it answers better.
The same question, to the same model:
- ✗ Before learning to draw below
- ✓ After learning to draw lower-right
Learning to draw ASCII improves spatial reasoning in language models.
What could learning to draw teach them?