Sketchpad enables certified exact arithmetic with a small vision-language model.
Sketchpad, a Qwen3.5-0.8B adapter, uses bounded-view checks and Lean proofs to certify exact arithmetic at any length, survive error, and compose.
Key Takeaways
- A Qwen3.5-0.8B model with the Sketchpad adapter is certified for exact arithmetic (+, −, ×, ÷, √, gcd) at any input length under a bounded-view setup.
- The certification checks 2,446 bounded views, with a finite proof reported in Lean, reducing all-length exactness to a finite set of fixed-size memory windows.
- Sketchpad scored 220/220 on pre-registered problems and 20/20 on 1,000-digit sums, outperforming tested frontier and open-weight models.
- The result remains robust to bounded numerical error and can be composed with other certified procedures, but it depends on the bounded-view condition.
- If steps may read earlier steps, no finite test suffices, so the guarantee does not apply to arbitrary small models or unrestricted access to prior outputs.
Meta description: Qwen3.5-0.8B Sketchpad adapter passes all 2,446 views, giving certified exact arithmetic for plus, minus, times, divide, square root, and gcd.
Tags: Exact Arithmetic, Lean, Qwen3.5-0.8B, Sketchpad, Formal Verification
Featured image: Abstract friendly illustration of a small model checking long digit sums, no real people, no logos, no text.
A Qwen3.5-0.8B model paired with a small adapter called Sketchpad passed all 2,446 bounded-view checks and is certified for exact arithmetic at any input length under the reported setup. The work also claims that the proof survives bounded numerical error and that certified procedures remain certified when combined.
The core claim: a finite proof for all lengths
The theory starts from a restriction: if each step of a model reads only a bounded view of an external memory, exactness on inputs of every length reduces to exactness on a finite set of views. A bounded view is a fixed-size window into stored input, and checking that window can cover an unbounded range of longer inputs. The author says this result was proved in Lean.
The report gives three extra theory points. First, necessity: once steps can also read earlier steps, no finite test suffices. Second, robustness: the guarantee survives bounded numerical error. Third, composition: certified procedures stay certified when combined.
What ran on a laptop and in the browser
Qwen3.5-0.8B plus a small adapter passes all 2,446 views on a laptop GPU, in the browser. The certified operations are addition, subtraction, multiplication, division, square root, and gcd at any length.
The result is presented as a checked procedure: the adapter plus the bounded-view rule is what allows the certification to apply across lengths, rather than only across the tested examples.
Scores on 220 pre-registered problems
Sketchpad (0.8B) scored 220. GPT-6.1 Sol scored 217, Claude Opus 5.5 scored 216, and Gemini 4 Argon scored 207. Five open-weight models scored between 149 and 199. Qwen3.5-27B scored 121, and the base Qwen3.5-0.8B scored 23.
On 1,000-digit sums, Sketchpad scored 20/20. The frontier models scored 19, 17, and 9. Every open-weight model scored 2 or fewer.
Why the result matters and where it stops
Exact arithmetic at every length can be treated as a provable property when the system is built around bounded views. The Qwen3.5-0.8B result suggests that exactness can be achieved with a small model when the arithmetic path is checked. It also shows why benchmark scores on named frontier models and open-weight models can look good while leaving gaps on very long digit sums.
The certification depends on the bounded-view condition. If a step can read earlier steps, the report says no finite test is enough. The guarantee is not a blanket promise that any small model, or any model with access to prior outputs, will be exact on all arithmetic inputs.
The author includes links to a live demo and code. The follow-on question is whether bounded-view certification can be extended to more operations or used as a design rule when building arithmetic tools around small models.

Comments (0)