NEWS
MARCH scales recurrent memory using content-routed state anchorsMARCH stores all past states in a memory bank and queries these anchors at each token with content-conditioned keys, increasing the memory capacity of fixed-size recurrent models in long contexts while keeping computational cost low; after pretraining, it consistently outperforms linear attention variants in joint reasoning, LongBench, and retrieval-in-context tasks. This is a news summary compiled for the AiPulse newsletter — not the original article. The original article is at arxiv.org. Read the original article AI Critique → English (automatic translation) |