
Author of "Why Stacking Sliding Windows Can't See Very Far" in Guangxuan Xiao's Blog
How media typically covers Guangxuan Xiao
Guangxuan Xiao as author
Stacking sliding window attention layers cannot achieve their theoretical context length due to information dilution and exponential barriers from residual connections, limiting effective memory to ~1,500 tokens despite theoretical capacity of 100,000+.
“Author of "Why Stacking Sliding Windows Can't See Very Far" in Guangxuan Xiao's Blog”