S. Baek

IWP: Token Pruning as Implicit Weight Pruning in Large Vision Language Models

This paper proposes a novel training free token pruning framework grounded in the dual form perspective of attention, and derives a novel metric quantifying both a token's information magnitude and information duplication.