Definition
Plain language
A system that makes a small AI do arithmetic one tiny column at a time, with no memory of earlier steps, so every possible step can be checked.
As stated in the literature
Mohsenvand's certified-arithmetic architecture: a frozen VLM plus adapter reads only an operation label and a three-cell rendered image, making the per-step input domain finite and exhaustively verifiable with logit-margin bounds.
Why it matters: By shrinking what the model sees at each step to a finite set of pictures, it makes it possible to check every situation the system can ever encounter rather than hoping test accuracy generalizes.
For example, to add two long numbers it shows the model just one column at a time, writes each answer digit onto an external sheet, and then slides over to the next column.
Heard on the show
“Today we're looking at Sketchpad, a preprint by Mostafa 'Neo' Mohsenvand at the MIT Media Lab.”Episode 291 — Certifying a Neural Network's Arithmetic by Taking Away Its Memory