Current LLMs struggle with complex, unsolved mathematical proofs that require deep logical reasoning. While synthetic data improves basic arithmetic, high-level theorem proving remains a bottleneck. Researchers are now testing specialized architectures to bridge this gap. Success here would shift AI from pattern matching to genuine discovery for mathematicians and physicists.