Tag
MathFormer is a small seq2seq model that achieves ~98.6% accuracy on symbolic math tasks, suggesting that mathematical reasoning in LLMs may be large-scale structured pattern completion rather than true reasoning.