Update README.md

README.md

@@ -98,6 +98,9 @@ wf = FlaxAutoModel.from_pretrained("nqs-models/j1j2_square_10x10", trust_remote_
 z = wf(wf.params, samples)
 ```
 
+Starting from the vector \\(z\\), a fully connected network can be trained to *fine-tune* the model on a different value of the ratio \\(J_2/J_1\\).
+See https://doi.org/10.1103/PhysRevResearch.6.023057 for more informations.
+
 Note: the hidden representation is well defined only for the non symmetrized model.
 
 #### Training Hyperparameters