Search papers, labs, and topics across Lattice.
2
0
6
0
Across reasoning, long-context understanding, and agentic tasks, SAS consistently outperforms trainable sparse attention baselines across attention budgets, with especially large gains under tight budgets, demonstrating more effective context ranking for downstream tasks.
Error accumulation in long video generation can be mitigated with a novel frequency-domain approach, achieving 24x extrapolation without the need for extensive training.
