Differentiable Efficient Operator Search — unified token reduction
arXiv paper unifies pruning, merging, pooling, and adaptive reweighting as regimes within a single operator space.
Proposes a differentiable search framework that jointly optimizes where to reduce tokens, how many to keep, and how to process the reduced information under budget constraints.
Recovers hand-designed operators as special cases, suggesting the framework captures the landscape of token-reduction strategies for multimodal models.