People say that adding code to the LLM training corpus improves the model's "reasoning" ability. If so, would including a training set of short programs followed by an execution trace improve its reasoning further as it would have to predict the executed code path from the src?

→ View original post on X — @replit