Small vision-language model achieves certified exact arithmetic
Original title: Sketchpad: certified exact arithmetic with a small vision-language model
Qwen3.5-0.8B with external sheet memory proves exact arithmetic for all input lengths after checking finite views in Lean.
Language models lack guarantees on long arithmetic problems. Sketchpad uses a small Qwen3.5-0.8B vision-language model that reads bounded three-cell views of an external working memory and emits arithmetic steps. All 2,446 views pass verification on a laptop GPU. The model solves signed addition, multiplication, division, square roots and GCDs at arbitrary length and outperforms GPT-6.1 Sol on pre-registered problems.
- Proven exact on all inputs via Lean theorem prover
- All 2,446 views verified on deployment hardware
- Solves 220 of 220 sums up to 1,000 digits
- Symbolic runtime computes polynomials exactly using model outputs
From the original article
1 min read at huggingface.co
Language models are accurate on most long-arithmetic problems, but their accuracy is estimated on samples and carries no guarantee for an unseen input; tool use is exact but moves the computation outside the network. We show that exactness at all input lengths becomes a finite, checkable property of a network when each of its steps reads only a bounded view of an external working memory. In Sketchpad, Qwen3.
By Mostafa 'Neo' Mohsenvand · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.