Search papers, labs, and topics across Lattice.
2
0
5
4
MemDreamer narrows the performance gap with human experts in long video understanding to just 3.7 points while processing only 2% of the full context.
Frontier models are wasted on routine GUI tasks: a step-level cascade that adaptively invokes stronger models only when lightweight monitors detect progress stalls or semantic drift slashes compute costs without sacrificing performance.
