PhD student, KAIST
1 paper at NeurIPS 2025
Token reduction with mitigating rank-collapsing (over-smoothing) through high-frequency token preservation