Raven: Linear-Time Sequence Model Achieves High-Recall via Sparse Memory Routing

bronzeagepapi · x · 2026-08-01

A new paper titled Raven: High-Recall Sequence Modeling with Sparse Memory Routing by Albert Gu and colleagues introduces Raven, a novel linear-time sequence model.

The model addresses the tradeoff in long-context recall: state-based linear models like SSMs write densely causing interference, while Sliding Window Attention (SWA) evicts old tokens entirely due to fixed windows. Raven maintains a fixed set of memory slots and, at each step, decays and updates only a selected subset via learned, input-dependent routing.

This approach mitigates SWA's hard eviction and SSMs' dense interference. Across recall-intensive benchmarks, Raven outperforms prior baselines and remains effective when extrapolating to context lengths up to 16x its training length.

Original post →

More from Research

Research channel →