arXivMasaaki Imaizumi, Masanori Koyama, Noboru Isobe, Kohei HayashiThu, May 28, 2026, 9:59 AM PDT
score 14.8
How Positional Encoding Prevents Transformer Attention Collapse
Original: Anti Mode-Collapse in Mean-Field Transformer via Auxiliary Variables
Source: arxiv.org ↗
Writing ELI5 summary…