Here is a study: https://journals.plos.org/plosone/article?id=10.1371/journal...
And, since LLMs are so bad at math currently, we may find that, by improving their math ability with gobs of synthetic data, we get improvements in general reasoning.
And, since LLMs are so bad at math currently, we may find that, by improving their math ability with gobs of synthetic data, we get improvements in general reasoning.
One aspect that's quite easy to criticise about this study is that it uses existing groups of students with different levels of maths training. This means that there is possibly self-selection etc, and one may argue there might also be a causal effect in the opposite direction (e.g. folks that are good at reasoning like to do maths.)