Search papers, labs, and topics across Lattice.
1
0
3
Transformers trained on a simple grid-world learn hidden representations that directly reflect the underlying predictive geometry, offering a glimpse into how neural networks internalize structural constraints.
