Researchers discovered a fundamental trade-off between gradient measurement efficiency and expressivity in quantum neural networks, proposing the Stabilizer-Logical Product Ansatz as an optimal solution that exploits quantum circuit symmetry to achieve efficient training while maintaining performance.