1 paper across 1 session
Token reduction with mitigating rank-collapsing (over-smoothing) through high-frequency token preservation