Jul 2026
Raven: High-Recall Sequence Modeling with Sparse Memory Routing
Interpolating between these models, Raven is introduced, a linear-time sequence model that maintains a fixed set of memory slots and, at each step, decays and updates only a selected subset via learned, input-dependent routing, thereby preserving long-range content much more effectively.
Arshia Afzal, Aviv Bick, Eric P. Xing et al.
· arXiv.org · 5 citations