In our new <a href="https://arxiv.org/abs/2502.02393">preprint</a>, we study how long chains of thought need to be for transformers to solve different algorithmic problems.