Search papers, labs, and topics across Lattice.
5
0
7
20
Layer patching can dramatically enhance model performance in size interpolation, revealing that simple strategies often outperform complex methods.
Uncovering hidden environmental factors can transform anti-poverty strategies, revealing actionable insights that were previously overlooked.
A new trust score enables reliable evaluation of conditional generation samples even when reference distributions are absent, leading to better quality and performance in real-world applications.
Discrete diffusion models can now be trained orders of magnitude faster by directly learning the underlying free-energy functional, bypassing the need for sample trajectory optimization.
Even with weaker assumptions, ICA post-processing can unlock state-of-the-art disentanglement from vanilla autoencoders and foundation model-scale MAEs.
