Progressive Memory Transformer adds window-aligned memory to transformers, enforcing local, mid-range, and global time-series objectives for strong low-label classification and forecasting.
Neural compressed sensing extends to function space to co-design wet-lab experiments with learning algorithms, achieving orders-of-magnitude higher information density by measuring multiple molecules simultaneously and deconvolving activity during training for antibodies and cell therapies.
MemDocAgent uses dependency-aware traversal and shared memory to generate consistent, hierarchical repository-level documentation in a single integrated context, outperforming both open- and closed-source baselines.