Neural networks often build in known symmetries, like the fact that a cat rotated is still a cat. But real data only roughly follows these rules, and strict symmetry can make the network hard to train, getting stuck or learning poorly.
The authors add a small, learnable "position signal" that temporarily breaks the symmetry during training, letting the network solve the problem more easily. They gradually turn this signal off, so at test time the network is back to being strictly symmetrical, but with better training results.
On several tasks including image classification and 3D point-cloud analysis, this method matches or improves accuracy compared to a strict symmetry model, while adding almost no extra computation or memory. The key finding is that a simple, nearly free trick can fix training problems caused by overly strict symmetry rules.