FG^2-GDN enhances long-context understanding with adaptive learning rates

Researchers have introduced FG$^2$-GDN, a novel approach to enhance long-context understanding in neural networks. This method improves upon existing Gated Delta Networks by replacing a scalar learning rate with a channel-wise vector, allowing for more dimension-specific adaptation. An extension, FG$^2$-GDN+, further refines control by decoupling scaling for keys and values, offering independent management of erasure and write strengths. Experiments indicate that these new variants achieve better associative recall and long-context comprehension with similar computational costs. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a new method for improving long-context understanding in neural networks, potentially impacting how models process and recall information over extended sequences.

RANK_REASON This is a research paper detailing a new method for enhancing neural network context understanding. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• FG^2-GDN
• Gated Delta Network
• Kimi Delta Attention
• GDN
• KDA
• arXiv
• AdaGrad
• Adam

• paper
• model release