Neural quantum state models can be steered to predict physical phenomena they were never trained on — and researchers now have a tool to see why.
A team studying neural quantum states (NQS) — a class of machine learning models used to approximate quantum many-body wavefunctions — applied sparse autoencoders to examine what these networks learn internally. Without any physical labels, the autoencoders extracted features from the models' internal activations that correlated strongly with measurable physical quantities: order parameters, staggered magnetization, and half-chain correlators. The finding holds across both static ground-state representations and real-time dynamics.
The deeper result is causal, not just correlational. By intervening on a single extracted feature post-training, the researchers could smoothly steer a corresponding physical observable — while leaving the model's variational energy nearly unchanged. That level of surgical control suggests NQS are encoding structured physical knowledge, not just curve-fitting to a loss function.
This matters because NQS have long been a black box: accurate, but opaque. Borrowing mechanistic interpretability methods from the large language model world and applying them to quantum physics models is a meaningful cross-domain transfer. Whether the approach scales to more complex quantum systems remains to be seen, but the early signal is that the physics is in there — researchers just needed the right scalpel to find it.