The Silent LaTeX Extraction Trap

Senior AI Engineer interview at Google DeepMind, and the interviewer asks:

“Your model crushes math word problems but falls apart the moment a user pastes LaTeX. A competitor’s model, with the same size and the same token budget, doesn’t have this problem. Where in your data pipeline do you look first?”

Don’t say: “We need more math data. Let’s fine-tune on arXiv or scale up.”