Proteus paper: incremental memory activation fixes linear attention's fixed-state saturation

behrouz_ali · x · 2026-09-15

A new paper thread introduces Proteus, addressing linear attention's core weakness at frontier scale: its fixed-size memory state saturates as context grows because all capacity is exposed from token 1, so early tokens waste capacity and later tokens inherit a full state. The proposed incremental memory activation schedules how much memory is live: early bottlenecks force summarization over memorization, then blocks unlock one at a time to admit new information. Proteus partitions memory into E blocks with gated reads/writes; the authors call it the simplest instantiation of the idea.

Original post →

More from Research

Research channel →