Search papers, labs, and topics across Lattice.
5
0
6
62
Learning to prune key-value caches based on future token utility allows KVpop to achieve up to 88% compression while retaining nearly full performance.
TiRex-2 achieves state-of-the-art performance while maintaining constant inference costs, revolutionizing how we approach multivariate time series forecasting.
xLSTM outperforms its competitors in complex sequence modeling tasks by leveraging advanced state tracking and memory correction mechanisms.
LLMs can match or exceed the reasoning performance of autoregressive methods, but with significantly improved compute efficiency, by using fixed-sequence "memory blocks" for latent reasoning.
xLSTM models can now effectively learn from large attention-based models, even outperforming their teachers on some tasks through a novel distillation and merging pipeline.
