Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
2
0
5
2
Recursive transformers can get a 43% training loss reduction by dynamically modulating LoRA weights based on the hidden state at each recurrence step, finally letting them do different things at different depths.
Object-centric world models can now disentangle continuous physics, discrete events, and abstract goals through a hierarchical latent space, achieving state-of-the-art next-state prediction.