I'm almost surprised that Gemini 3 uniquely has this problem. I would have expected that responses from any LLM that require complex math notation would almost certainly be LaTeX heavy, given the abundance of LaTeX source material in the training data. I suppose it is a flaw if a model can't avoid LaTeX, but given that it is the standard (and for the foreseeable future too) I don't know what appropriate output would look like. For "pure" mathematics or similar topics I think LaTeX (or system that represents a superset of LaTeX) is the only acceptable option.