Jin Hyeon Kim

PhD student, Korea Advanced Institute of Science & Technology

1 paper at NeurIPS 2025

Homepage· OpenReview· Semantic Scholar· Google Scholar

Poster Session 6

Friday, December 5, 2025 · 4:30 PM → 7:30 PM

First Attentions Last: Better Exploiting First Attentions for Efficient Parallel Training

#810 · Gyudong Kim, Hyukju Na, Jin Hyeon Kim, Hyunsung Jang, Jaemin Park, Jaegi Hwang, NAMKOO HA, Seungryong Kim, Young Geun Kim

New Transformer architectures, FAL and FAL+, reuse the first attention output carefully —FAL cuts TP communication for up to 44% faster training, while FAL+ improves perplexity.